Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Dynamic Games with Almost Perfect Information
Wei He∗ Yeneng Sun†
This version: October 8, 2015
Job Market Paper
Abstract
This paper aims to solve two fundamental problems on finite or infinite horizon
dynamic games with complete information. Under some mild conditions, we
prove (1) the existence of subgame-perfect equilibria in general dynamic games
with simultaneous moves (i.e. almost perfect information), and (2) the existence
of pure-strategy subgame-perfect equilibria in perfect-information dynamic games
with uncertainty. Our results go beyond previous works on continuous dynamic
games in the sense that public randomization and the continuity requirement on the
state variables are not needed. As an illustrative application, a dynamic stochastic
oligopoly market with intertemporally dependent payoffs is considered.
∗Department of Economics, The University of Iowa, W249 Pappajohn Business Building, Iowa City,
IA 52242. E-mail: [email protected].†Department of Economics, National University of Singapore, 1 Arts Link, Singapore 117570. Email:
1
Contents
1 Introduction 3
2 Model and main result 6
3 Dynamic oligopoly market with sticky prices 9
4 Variations of the main result 10
4.1 Dynamic games with partially perfect information and a generalized
ARM condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4.2 Continuous dynamic games with partially perfect information . . . 13
5 Appendix 15
5.1 Technical preparations . . . . . . . . . . . . . . . . . . . . . . . . . 16
5.2 Discontinuous games with endogenous stochastic sharing rules . . . 33
5.3 Proofs of Theorem 1 and Proposition 1 . . . . . . . . . . . . . . . . 37
5.3.1 Backward induction . . . . . . . . . . . . . . . . . . . . . . 37
5.3.2 Forward induction . . . . . . . . . . . . . . . . . . . . . . . 37
5.3.3 Infinite horizon case . . . . . . . . . . . . . . . . . . . . . . 42
5.4 Proof of Proposition 2 . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.5 Proof of Proposition 3 . . . . . . . . . . . . . . . . . . . . . . . . . 61
References 63
2
1 Introduction
Dynamic games with complete information, where players observe the history and
move simultaneously or alternatively in finite or infinite horizon, arise naturally in
many situations. As noted in Fudenberg and Tirole (1991, Chapter 4) and Harris,
Reny and Robson (1995), these games form a general class of regular dynamic
games with many applications in economics, political science, and biology. The
associated notion of subgame-perfect equilibrium is a fundamental game-theoretic
concept. For games with finitely many actions and stages, Selten (1965) showed
the existence of subgame-perfect equilibria. The infinite horizon but finite-action
case was covered by Fudenberg and Levine (1983).
Since the agents in many economic models need to make continuous choices,
it is important to consider dynamic games with general action spaces. However,
Harris, Reny and Robson (1995) presented a simple example without subgame-
perfect equilibrium, where the game has two players in each of the two stages with
only one player having a continuous choice set.1 Thus, the existence of subgame-
perfect equilibria under some suitable conditions remains an open problem even
for two-stage dynamic games. The purpose of this paper is to prove the existence
of subgame-perfect equilibria for general dynamic games with simultaneous or
alternative moves in finite or infinite horizon under some suitable conditions.
We shall adopt the general forms of intertemporal preferences, requiring neither
stationarity nor additive separability.
For deterministic games with perfect information (i.e., the players move
alternatively), the existence of pure-strategy subgame-perfect equilibria was shown
in Harris (1985), Hellwig and Leininger (1987), Borgers (1989, 1991) and Hellwig et
al. (1990) with the model parameters being continuous in actions. However, if the
“deterministic” assumption is dropped by introducing a passive player - Nature,
then pure-strategy subgame-perfect equilibrium need not exist as shown by a four-
stage game in Harris, Reny and Robson (1995, p. 538). In fact, Luttmer and
Mariotti (2003) even demonstrated the nonexistence of mixed-strategy subgame-
perfect equilibrium in a five-stage game.2 Thus, it is still an open problem to show
the existence of (pure or mixed-strategy) subgame-perfect equilibria in (finite or
infinite horizon) perfect-information dynamic games with uncertainty under some
general conditions.
1As noted in Section 2 of Harris, Reny and Robson (1995), such a dynamic game with continuouschoices provides a minimal nontrivial counterexample; see also Exercise 13.4 in Fudenberg and Tirole(1991) for another counterexample of Harris.
2All those counterexamples show that various issues arise when one considers general dynamic games.In the setting with incomplete information, even the equilibrium notion needs to be carefully treated;see Myerson and Reny (2015).
3
Harris, Reny and Robson (1995) considered continuous dynamic games with
almost perfect information in the sense that the players move simultaneously.
In such games, there are a finite number of active players and a passive player
(Nature), and all the relevant model parameters are assumed to be continuous in
both action and state variables (i.e., Nature’s moves). Harris, Reny and Robson
(1995) showed the existence of subgame-perfect correlated equilibria by introducing
a public randomization device.3 As mentioned above, they also demonstrated the
possible nonexistence of subgame-perfect equilibrium through a simple example.
This paper resolves the above two open problems (in both finite and infinite
horizon) not just for the class of continuous dynamic games as considered in
the earlier works, but more importantly, for general dynamic games in which
the relevant model parameters are assumed to be continuous in actions and
measurable in states.4 In particular, we show the existence of a subgame-
perfect equilibrium in a general dynamic game with almost perfect information
under some suitable conditions on the state transitions. Theorem 1 (and also
Proposition 2) below goes beyond earlier works on continuous dynamic games
by dropping public randomization and the continuity requirement on the state
variables. Thus, the class of games considered here includes general stochastic
games, where the stage payoffs are usually assumed to be continuous in actions and
measurable in states.5 Our Proposition 1 in Section 2 also presents some regularity
properties of the equilibrium payoff correspondences, including compactness and
upper hemicontinuity in the action variables.6 As an illustrative application of
Theorem 1, we consider a dynamic oligopoly market in which firms face stochastic
demand/cost and intertemporally dependent payoffs.
We work with the condition that the state transition in each period (except
for those periods with one active player) has a component with a suitable density
function with respect to some atomless reference measure. This condition is also
minimal in the particular sense that the existence result may fail to hold if (1) the
passive player, Nature, is not present in the model as shown in Harris, Reny and
3See also Mariotti (2000) and Reny and Robson (2002).4Since the agents need to make optimal choices, the continuity assumption in terms of actions
is natural and widely adopted. However, the state variable is not a choice variable, and thus itis unnecessary to impose the state continuity requirement in a general model. Note that the statemeasurability assumption is the minimal regularity condition one would expect for the model parameters.In addition, we point out that the proof of our main result in the case with state continuity as inSubsection 5.5 is much simpler than that of the general case.
5Proposition 2 below does imply a new existence result on subgame-perfect equilibria for a generalstochastic game; see Remark 2 in Subsection 4.1.
6Such an upper hemicontinuity property in terms of correspondences of equilibrium payoffs, oroutcomes, or correlated strategies has been the key for proving the relevant existence results as inHarris (1985), Hellwig and Leininger (1987), Borgers (1989, 1991), Hellwig et al. (1990), Harris, Renyand Robson (1995) and Mariotti (2000).
4
Robson (1995), or (2) with the presence of Nature, the reference measure is not
atomless as shown in Luttmer and Mariotti (2003).
For the special class of continuous dynamic games with almost perfect
information, we can weaken the atomless reference measure condition slightly.
In particular, we simply assume the state transition in each period (except for
those periods with one active player) to be an atomless probability measure
for any given history, without the requirement of a common reference measure.
Note that our model allows the state history to fully influence all the model
parameters. Clearly, it covers the case with public randomization in the sense that
the state transition has an additional atomless component that is independently
and identically distributed across time, and does not enter the payoffs, state
transitions and action correspondences. Thus, we obtain the existence result in
Harris, Reny and Robson (1995) as a special case.
For dynamic games with almost perfect information, our main result allows
the players to take mixed strategies. However, for the special class of dynamic
games with perfect information,7 we obtain the existence of pure-strategy subgame-
perfect equilibria in Corollaries 2 and 3. When Nature is present, the only known
general existence result for dynamic games with perfect information is, to the
best of our knowledge, for continuous games with public randomization. On the
contrary, our Corollary 2 needs neither continuity in the state variables nor public
randomization. Furthermore, our Corollary 3 provides a new existence result for
continuous dynamic games with perfect information, which generalizes the results
of Harris (1985), Hellwig and Leininger (1987), Borgers (1989), and Hellwig et al.
(1990) to the case when Nature is present.
To prove Theorem 1, we follow the standard three-step procedure in obtaining
subgame-perfect equilibria of dynamic games, namely, backward induction, forward
induction, and approximation of infinite horizon by finite horizon. Because we
drop public randomization and the continuity requirement on the state variables,
new technical difficulties arise in each step of the proof. In the step of backward
induction, we obtain a new existence result for discontinuous games with stochastic
endogenous sharing rules, which extends the main result of Simon and Zame
(1990) by allowing the payoff correspondence to be measurable (instead of upper
hemicontinuous) in states. For forward induction, we need to obtain strategies
that are jointly measurable in history. When there is a public randomization
device, the joint measurability follows from the measurable version of Skorokhod’s
representation theorem and implicit function theorem respectively as in Harris,
7Dynamic games with perfect information do have wide applications. For some examples, see Phelpsand Pollak (1968) for an intergenerational bequest game, and Peleg and Yaari (1973) and Goldman(1980) for intrapersonal games in which consumers have changing preferences.
5
Reny and Robson (1995) and Reny and Robson (2002). Here we need to work with
the deep “measurable” measurable choice theorem of Mertens (2003). Our main
existence result for the finite-horizon case then follows immediately from backward
and forward inductions. Lastly, in order to extend the result to the infinite horizon
setting, we need to handle various subtle measurability issues due to the lack of
continuity in the state variables in our model, which is the most difficult part of
the proof for Theorem 1.8 As noted in Subsection 5.5 below, the proof for the
existence result in the case of continuous dynamic games (i.e., Proposition 3) is
much simpler than that of Theorem 1.
The rest of the paper is organized as follows. The model and main result are
presented in Section 2. An illustrative application of Theorem 1 to a dynamic
oligopoly market with stochastic demand/cost and intertemporally dependent
payoffs is given in Section 3. Section 4 provides several variations of the main
result. All the proofs are left in the Appendix.
2 Model and main result
In this section, we shall present the model for an infinite-horizon dynamic game
with almost perfect information.
The set of players is I0 = {0, 1, . . . , n}, where the players in I = {1, . . . , n} are
active and player 0 is Nature. All players move simultaneously. Time is discrete,
and indexed by t = 0, 1, 2, . . ..
The set of starting points is a closed set H0 = X0×S0, where X0 is a compact
metric space and S0 is a Polish space (that is, a complete separable metric space).9
At stage t ≥ 1, player i’s action will be chosen from a subset of a Polish space Xti
for each i ∈ I, and Xt =∏i∈I Xti. Nature’s action is chosen from a Polish space
St. Let Xt =∏
0≤k≤tXk and St =∏
0≤k≤t Sk. The Borel σ-algebras on Xt and
St are denoted by B(Xt) and B(St), respectively. Given t ≥ 0, a history up to the
stage t is a vector
ht = (x0, s0, x1, s1, . . . , xt, st) ∈ Xt × St.
The set of all such possible histories is denoted by Ht. For any t ≥ 0, Ht ⊆ Xt×St.For any t ≥ 1 and i ∈ I, let Ati be a measurable, nonempty and compact
valued correspondence from Ht−1 to Xti such that (1) Ati is sectionally continuous
8Because our relevant model parameters are only measurable in the state variables, the usual methodof approximating a limit continuous dynamic game by a sequence of finite games, as used in Hellwig etal. (1990), Borgers (1991) and Harris, Reny and Robson (1995), is not applicable in our setting.
9In each stage t ≥ 1, there will be a set of action profiles Xt and a set of states St. Without loss ofgenerality, we assume that the set of initial points is also a product space for notational consistency.
6
on Xt−1,10 and (2) Ati(ht−1) is the set of available actions for player i ∈ I given
the history ht−1.11 Let At =∏i∈I Ati. Then Ht = Gr(At) × St, where Gr(At) is
the graph of At.
For any x = (x0, x1, . . .) ∈ X∞, let xt = (x0, . . . , xt) ∈ Xt be the truncation of
x up to the period t. Truncations for s ∈ S∞ can be defined similarly. Let H∞ be
the subset of X∞ × S∞ such that (x, s) ∈ H∞ if (xt, st) ∈ Ht for any t ≥ 0. Then
H∞ is the set of all possible histories in the game.12
For any t ≥ 1, Nature’s action is given by a Borel measurable mapping ft0 from
Ht−1 to M(St) such that ft0 is sectionally continuous on Xt−1, where M(St) is
endowed with the topology induced by the weak convergence.13 For each t ≥ 0,
suppose that λt is a Borel probability measure on St and λt is atomless for t ≥ 1.
Let λt = ⊗0≤k≤tλt for t ≥ 0. We shall assume the following condition on the state
transitions.
Assumption 1 (Atomless Reference Measure (ARM)). A dynamic game is said
to satisfy the “atomless reference measure (ARM)” condition if for each t ≥ 1,
1. the probability ft0(·|ht−1) is absolutely continuous with respect to λt on St
with the Radon-Nikodym derivative ϕt0(ht−1, st) for all ht−1 ∈ Ht−1;14
2. the mapping ϕt0 is Borel measurable and sectionally continuous on Xt−1, and
integrably bounded in the sense that there is a λt-integrable function φt : St →R+ such that ϕt0(ht−1, st) ≤ φt(st) for any ht−1 ∈ Ht−1 and st ∈ St.
For each i ∈ I, the payoff function ui is a Borel measurable mapping from H∞
to R++ which is sectionally continuous on X∞ and bounded by γ > 0.15
When one considers a dynamic game with infinite horizon, the following
“continuity at infinity” condition is standard.16 In particular, all discounted
repeated games or stochastic games satisfy this condition.
10Suppose that Y1, Y2 and Y3 are all Polish spaces, and Z ⊆ Y1 × Y2. Denote Z(y1) = {y2 ∈Y2 : (y1, y2) ∈ Z} for any y1 ∈ Y1. A function (resp. correspondence) f : Z → Y3 is said to be sectionallycontinuous on Y2 if f(y1, ·) is continuous on Z(y1) for all y1 with Z(y1) 6= ∅. Similarly, one can definethe sectional upper hemicontinuity for a correspondence.
11Suppose that Y and Z are both Polish spaces, and Ψ is a correspondence from Y to Z. Hereafter,the measurability of Ψ, unless specifically indicated, is with respect to the Borel σ-algebra B(Y ) on Y .
12A finite horizon dynamic game can be regarded as a special case of an infinite horizon dynamicgame in the sense that the action correspondence Ati is point-valued for each player i ∈ I and t ≥ T forsome stage T ≥ 1; see, for example, Borgers (1989) and Harris, Reny and Robson (1995).
13For a Polish space A, M(A) denotes the set of all Borel probability measures on A, and 4(A) isthe set of all finite Borel measures on A.
14It is common to have a reference measure when one considers a game with uncountable states. Forexample, if St is a subset of Rl, then the Lebesgue measure is a natural reference measure.
15Since ui is bounded, we can assume that the value of the payoff function is strictly positive withoutloss of generality.
16See, for example, Fudenberg and Levine (1983).
7
For any T ≥ 1, let
wT = supi∈I
(x,s)∈H∞(x,s)∈H∞xT−1=xT−1
sT−1=sT−1
|ui(x, s)− ui(x, s)|. (1)
Assumption 2 (Continuity at Infinity). A dynamic game is said to be “continuous
at infinity” if wT → 0 as T →∞.
For player i ∈ I, a strategy fi is a sequence {fti}t≥1 such that fti is a Borel
measurable mapping from Ht−1 to M(Xti) with fti(Ati(ht−1)|ht−1) = 1 for all
ht−1 ∈ Ht−1. A strategy profile f = {fi}i∈I is a combination of strategies of all
active players.
In any subgame, a strategy combination will generate a probability distribution
over the set of possible histories. This probability distribution is called the path
induced by the strategy combination in this subgame.
Definition 1. Suppose that a strategy profile f = {fi}i∈I and a history ht ∈ Ht are
given for some t ≥ 0. Let τt = δht, where δht is the probability measure concentrated
at the one point ht. If τt′ ∈M(Ht′) has already been defined for some t′ ≥ t, then
let
τt′+1 = τt′ � (⊗i∈I0f(t′+1)i).17
Finally, let τ ∈ M(H∞) be the unique probability measure on H∞ such that
MargHt′τ = τt′ for all t′ ≥ t. Then τ is called the path induced by f in the
subgame ht. For all i ∈ I,∫H∞
ui dτ is the payoff of player i in this subgame.
The notion of subgame-perfect equilibrium is defined as follows.
Definition 2 (SPE). A subgame-perfect equilibrium is a strategy profile f such
that for all i ∈ I, t ≥ 0, and λt-almost all ht ∈ Ht,18 player i cannot improve his
payoff in the subgame ht by a unilateral change in his strategy.19
17Denote ⊗i∈I0f(t′+1)i as a transition probability from Ht′ to M(Xt′+1). Notice that the strategyprofile is usually represented by a vector. For the notational simplicity later on, we assume that⊗i∈I0f(t′+1)i(·|ht′) represents the strategy profile in stage t′ + 1 for a given history ht′ ∈ Ht′ , where⊗i∈I0f(t′+1)i(·|ht′) is the product of the probability measures f(t′+1)i(·|ht′), i ∈ I0. If λ is a finitemeasure on X and ν is a transition probability from X to Y , then λ � ν is a measure on X × Y suchthat λ � ν(A×B) =
∫Aν(B|x)λ(dx) for any measurable subsets A ⊆ X and B ⊆ Y .
18A property is said to hold for λt-almost all ht = (xt, st) ∈ Ht if it is satisfied for λt-almost all st ∈ Stand all xt ∈ Ht(s
t).19When the state space is uncountable and has a reference measure, it is natural to consider the
optimality for almost all sub-histories in the probabilistic sense; see, for example, Abreu, Pearce andStacchetti (1990) and Footnote 4 therein.
8
The following theorem is our main result, which shows the existence of a
subgame-perfect equilibrium under the conditions of ARM and continuity at
infinity. Its proof is left in the appendix.
Theorem 1. If a dynamic game satisfies the ARM condition and is continuous at
infinity, then it possesses a subgame-perfect equilibrium.
Let Et(ht−1) be the set of subgame-perfect equilibrium payoffs in the subgame
ht−1. The following result demonstrates the compactness and upper hemicontinuity
properties of the correspondence Et.
Proposition 1. If a dynamic game satisfies the ARM condition and is continuous
at infinity, then Et is nonempty and compact valued, and essentially sectionally
upper hemicontinuous on Xt−1.20
3 Dynamic oligopoly market with sticky prices
In this section, we consider a dynamic oligopoly market in which firms face
stochastic demand/cost and intertemporally dependent payoffs. Such a model is
a variant of the well-known dynamic oligopoly models as considered in Green and
Porter (1984) and Rotemberg and Saloner (1986), which examined the response of
firms for demand fluctuations. The key feature of our example is the existence of
sticky price effect, which means that the desirability of the good from the demand
side could depend on the accumulated past output, and hence gives intertemporally
dependent payoff functions.
We consider a dynamic oligopoly market in which n firms produce a homoge-
neous good in an infinite-horizon setting. The inverse demand function is denoted
by Pt(Q1, . . . , Qt, st), whereQt is the industry output and st the observable demand
shock in period t. Notice that the price depends on the past outputs. One possible
reason could be that the desirability of consumers will be influenced by their
previous consumptions, and hence the price does not adjust instantaneously. We
assume that Pt is a bounded function which is continuous in (Q1, . . . , Qt) and
measurable in st. In period t, the shock st is selected from the set St = [at, bt].
We denote firm i’s output in period t by qti so that Qt =∑n
i=1 qti. The cost of
firm i in period t is cti(qti, st) given the output qti and the shock st, where cti is a
bounded function continuous in qti and measurable in st. The discount factor of
firm i is βi ∈ [0, 1).
20Suppose that Y1, Y2 and Y3 are all Polish spaces, and Z ⊆ Y1×Y2 and η is a Borel probability measureon Y1. Denote Z(y1) = {y2 ∈ Y2 : (y1, y2) ∈ Z} for any y1 ∈ Y1. A function (resp. correspondence)f : Z → Y3 is said to be essentially sectionally continuous on Y2 if f(y1, ·) is continuous on Z(y1) for η-almost all y1. Similarly, one can define the essential sectional upper hemicontinuity for a correspondence.
9
The timing of events is as follows.
1. At the beginning of period t, all firms learn the realization of st, which
is determined by the law of motion κt(·|s1, Q1, . . . , st−1, Qt−1). Suppose
that κt(·|s1, Q1, . . . , st−1, Qt−1) is absolutely continuous with respect to the
uniform distribution on St with density ϕt(s1, Q1, . . . , st−1, Qt−1, st), where
ϕt is bounded, continuous in (Q1, . . . , Qt−1) and measurable in (s1, . . . , st).
2. Firms then simultaneously choose the level of their output qt = (qt1, . . . , qtn),
where qti ∈ Ati(st, Qt−1) ⊆ Rl for i = 1, 2, . . . , n. In particular, the
correspondence Ati gives the available actions of firm i, which is nonempty
and compact valued, measurable in st, and continuous in Qt−1.
3. The strategic choices of all the firms then become common knowledge and
this one-period game is repeated.
In period t, given the shock st and the output {qk}1≤k≤t up to time t with
qk = (qk1, . . . , qkn), the payoff of firm i is
uti(q1, . . . , qt, st) =
Pt( n∑j=1
q1j , . . . ,
n∑j=1
qtj , st)− cti(qti, st)
qti.Given a sequence of outputs {qt}t≥1 and shocks {st}t≥1, firm i receives the payoff
u1i(q1, s1) +
∞∑t=2
βt−1i uti(q1, . . . , qt, st).
Remark 1. Our dynamic oligopoly model has a non-stationary structure. In
particular, the transitions and payoffs are history-dependent. The example
captures the scenario that the price of the homogeneous product does not adjust
instantaneously to the price indicated by its demand function at the given level
of output. For more applications with intertemporally dependent utilities, see, for
example, Ryder and Heal (1973), Fershtman and Kamien (1987) and Becker and
Murphy (1988). If the model is stationary and the inverse demand function only
depends the current output, then the example reduces to be the dynamic oligopoly
game with demand fluctuations as considered in Rotemberg and Saloner (1986).
By condition (1) above, the ARM condition is satisfied. It is also easy to see
that the game is continuous at infinity. By Theorem 1, we have the following result.
Corollary 1. The dynamic oligopoly market possesses a subgame-perfect equilib-
rium.
10
4 Variations of the main result
In this section, we will consider several variations of our main result.
In Subsection 4.1, we still consider dynamic games whose parameters are
continuous in actions and measurable in states. We partially relax the ARM
condition in two ways. First, we allow the possibility that there is only one
active player (but no Nature) at some stages, where the ARM type condition
is dropped. Second, we introduce an additional weakly continuous component on
the state transitions at any other stages. In addition, we allow the state transition
in each period to depend on the current actions as well as on the previous history.
Thus, we combine the models for dynamic games with perfect and almost perfect
information. We show the existence of a subgame perfect equilibrium such that
whenever there is only one active player at some stage, the player can play pure
strategy as part of the equilibrium strategies. As a byproduct, we obtain a new
existence result for stochastic games. The existence of pure-strategy subgame-
perfect equilibria for dynamic games with perfect information (with or without
Nature) is provided as an immediate corollary.
In Subsection 4.2, we consider the special case of continuous dynamic games
in the sense that all the model parameters are continuous in both action and
state variables. We can obtain the corresponding results under a slightly weaker
condition. All the previous existence results for continuous dynamic games with
perfect and almost perfect information are covered as our special cases.
We will follow the setting and notations in Section 2 as closely as possible. For
simplicity, we only describe the changes we need to make on the model. All the
proofs are left in the appendix.
4.1 Dynamic games with partially perfect information
and a generalized ARM condition
In this subsection, we will generalize the model in Section 2 in three directions.
The ARM condition is partially relaxed such that (1) perfect information may be
allowed in some stages, and (2) the state transitions have a weakly continuous
component in all other stages. In addition, the state transition in any period can
depend on the action profile in the current stage as well as on the previous history.
The fist change allows us to combine the models of dynamic games with perfect
and almost perfect information. The second generalization implies that the state
transitions need not be norm continuous on the Banach space of finite measures.
The last modification covers the model of stochastic games as a special case.
The changes are described below.
11
1. The state space is a product space of two Polish spaces; that is, St = St × Stfor each t ≥ 1.
2. For each i ∈ I, the action correspondence Ati from Ht−1 to Xti is measurable,
nonempty and compact valued, and sectionally continuous on Xt−1 × St−1.
The additional component of Nature is given by a measurable, nonempty
and closed valued correspondence At0 from Gr(At) to St, which is sectionally
continuous on Xt × St−1. Then Ht = Gr(At0)× St, and H∞ is the subset of
X∞ × S∞ such that (x, s) ∈ H∞ if (xt, st) ∈ Ht for any t ≥ 0.
3. The choices of Nature depend not only on the history ht−1, but also on the
action profile xt in the current stage. The state transition ft0(ht−1, xt) =
ft0(ht−1, xt) � ft0(ht−1, xt), where ft0 is a transition probability from Gr(At)
toM(St) such that ft0(At0(ht−1, xt)|ht−1, xt) = 1 for all (ht−1, xt) ∈ Gr(At),
and ft0 is a transition probability from Gr(At0) to M(St).
4. For each i ∈ I, the payoff function ui is a Borel measurable mapping from H∞
to R++ which is bounded by γ > 0, and sectionally continuous on X∞× S∞.
We allow the possibility for the players to have perfect information in some
stages. For t ≥ 1, let
Nt =
1, if ft0(ht−1, xt) ≡ δst for some st and
|{i ∈ I : Ati is not point valued}| = 1;
0, otherwise,
where |K| represents the number of points in the set K. Thus, if Nt = 1 for some
stage t, then the player who is active in the period t is the only active player and
has perfect information.
We will drop the ARM condition in those periods with only one active player,
and weaken the ARM condition in other periods.
Assumption 3 (ARM′). 1. For any t ≥ 1 with Nt = 1, St is a singleton set
{st} and λt = δst.
2. For each t ≥ 1 with Nt = 0, ft0 is sectionally continuous on Xt × St−1,
where the range space M(St) is endowed with topology of weak convergence
of measures on St. The probability measure ft0(·|ht−1, xt, st) is absolutely
continuous with respect to an atomless Borel probability measure λt on St
for all (ht−1, xt, st) ∈ Gr(At0), and ϕt0(ht−1, xt, st, st) is the corresponding
density.21
21In this subsection, a property is said to hold for λt-almost all ht ∈ Ht if it is satisfied for λt-almostall st ∈ St and all (xt, st) ∈ Ht(s
t).
12
3. The mapping ϕt0 is Borel measurable and sectionally continuous on Xt ×St, and integrably bounded in the sense that there is a λt-integrable function
φt : St → R+ such that ϕt0(ht−1, xt, st, st, ) ≤ φt(st) for any (ht−1, xt, st).
The following proposition shows that the existence result is still true in this
more general setting.
Proposition 2. If an infinite-horizon dynamic game satisfies the ARM′ condition
and is continuous at infinity, then it possesses a subgame-perfect equilibrium f . In
particular, for j ∈ I and t ≥ 1 such that Nt = 1 and player j is the only active
player in this period, ftj can be deterministic. Furthermore, the equilibrium payoff
correspondence Et is nonempty and compact valued, and essentially sectionally
upper hemicontinuous on Xt−1 × St−1.
Remark 2. The proposition above also implies a new existence result of subgame-
perfect equilibria for stochastic games. Consider a standard stochastic game
with uncountable states as in Mertens and Parthasarathy (2003). Mertens and
Parthasarathy (2003) proved the existence of a subgame-perfect equilibrium by
assuming the state transitions to be norm continuous with respect to the actions in
the previous stage. On the contrary, our Proposition 2 allows the state transitions
to have a weakly continuous component.
Dynamic games with perfect information is a special class of dynamic games
in which players move sequentially. As noted in Footnote 7, such games have
been extensively studied and found wide applications in economics. As an
immediate corollary, an equilibrium existence result for dynamic games with perfect
information is given below.
Corollary 2. If a dynamic game with perfect information satisfies the ARM′
condition and is continuous at infinity, then it possesses a pure-strategy subgame-
perfect equilibrium.
4.2 Continuous dynamic games with partially perfect
information
In this subsection, we will study an infinite-horizon dynamic game with a
continuous structure. As in the previous subsection, we allow the state transition
to depend on the action profile in the current stage as well as on the previous
history, and the players may have perfect information in some stages.
1. For each t ≥ 1, the choices of Nature depends not only on the history ht−1, but
also on the action profile xt in this stage. For any t ≥ 1, suppose that At0 is a
13
continuous, nonempty and closed valued correspondence from Gr(At) to St.
Then Ht = Gr(At0), and H∞ is the subset of X∞×S∞ such that (x, s) ∈ H∞if (xt, st) ∈ Ht for any t ≥ 0.
2. Nature’s action is given by a mapping ft0 from Gr(At) to M(St) such that
ft0(At0(ht−1, xt)|ht−1, xt) = 1 for all (ht−1, xt) ∈ Gr(At).
3. For each t ≥ 1, let
Nt =
1, if ft0(ht−1, xt) ≡ δst for some st and
|{i ∈ I : Ati is not point valued}| = 1;
0, otherwise.
Definition 3. A dynamic game is said to be continuous if for each t and i,
1. the action correspondence Ati is continuous on Ht−1;
2. the transition probability ft0 is a continuous mapping from Gr(At) to M(St),
where M(St) is endowed with topology of weak convergence of measures on
St;
3. the payoff function ui is continuous on H∞.
Note that the “continuity at infinity” condition is automatically satisfied in a
continuous dynamic game.
Next, we propose the condition of “atomless transitions” on the state space,
which means that the state transition is an atomless probability measure in any
stage. This condition is slightly weaker than the ARM condition.
Assumption 4 (Atomless Transitions). 1. For any t ≥ 1 with Nt = 1, St is a
singleton set {st}.
2. For each t ≥ 1 with Nt = 0, ft0(ht−1) is an atomless Borel probability measure
for each ht−1 ∈ Ht−1.
Since we work with continuous dynamic games, we can adopt a slightly stronger
notion of subgame-perfect equilibrium. That is, each player’s strategy is optimal
in every subgame given the strategies of all other players.
Definition 4 (SPE′). A subgame-perfect equilibrium is a strategy profile f such
that for all i ∈ I, t ≥ 0, and all ht ∈ Ht, player i cannot improve his payoff in the
subgame ht by a unilateral change in his strategy.
The result on the equilibrium existence is presented below.
14
Proposition 3. If a continuous dynamic game has atomless transitions, then it
possesses a subgame-perfect equilibrium f . In particular, for j ∈ I and t ≥ 1
such that Nt = 1 and player j is the only active player in this period, ftj can
be deterministic. In addition, Et is nonempty and compact valued, and upper
hemicontinuous on Ht−1 for any t ≥ 1.
Remark 3. Proposition 3 goes beyond the main result of Harris, Reny and Robson
(1995). They proved the existence of a subgame-perfect correlated equilibrium
in a continuous dynamic game with almost perfect information by introducing a
public randomization device, which does not influence the payoffs, transitions or
action correspondences. It is easy to see that their model automatically satisfies the
condition of atomless transitions. The state in our model is completely endogenous
in the sense that it affects all the model parameters such as payoffs, transitions,
and action correspondences.
Remark 4. Proposition 3 above provides a new existence result for continuous
stochastic games. As remarked in the previous subsection, the existence of subgame-
perfect equilibria has been proved for general stochastic games with a stronger
continuity assumption on the state transitions, namely the norm continuity. On
the contrary, we only need to require the state transitions to be weakly continuous.
Remark 5. The condition of atomless transitions is minimal. In particular, the
counterexample provided by Luttmer and Mariotti (2003), which is a continuous
dynamic game with perfect information and Nature, does not have any subgame-
perfect equilibrium. In their example, Nature is active in the third period, but the
state transitions could have atoms. Thus, our condition of atomless transitions is
violated.
The next corollary follows from Proposition 3, which presents the existence
result for continuous dynamic games with perfect information (and Nature).
Corollary 3. If a continuous dynamic game with perfect information has atomless
transitions, then it possesses a pure-strategy subgame-perfect equilibrium.
Remark 6. Harris (1985), Hellwig and Leininger (1987), Borgers (1989) and
Hellwig et al. (1990) proved the existence of subgame-perfect equilibria in contin-
uous dynamic games with perfect information. In particular, Nature is absent in
all those papers. Luttmer and Mariotti (2003) provided an example of a five-stage
continuous dynamic game with perfect information, in which Nature is present
and no subgame-perfect equilibrium exists. The only known general existence
result, to the best of our knowledge, for (finite or infinite horizon) continuous
dynamic games with perfect information and Nature is the existence of subgame-
perfect correlated equilibria via public randomization as in Harris, Reny and Robson
(1995). Corollary 3 covers all those existence results as special cases.
15
5 Appendix
In Subsection 5.1, we present Lemmas 1-15 as the mathematical preparations for
proving Theorem 1. Since correspondences will be used extensively in the proofs,
we collect, for the convenience of the reader, several known results on various
properties of correspondences as Lemmas 1-7. The proof for the finite-horizon case
of Theorem 1 will depend on Lemmas 1-7. The rest of the lemmas (Lemmas 8-15)
will only be used for proving the infinite-horizon case.
We present in Subsection 5.2 a new existence result for discontinuous games
with stochastic endogenous sharing rules. The proofs of Theorem 1 and Proposition
1 are given in Subsection 5.3. Finally, the proofs for the results in Section 4 are
provided in Subsections 5.4 and 5.5.
5.1 Technical preparations
Let (S,S) be a measurable space and X a topological space with its Borel σ-algebra
B(X). A correspondence Ψ from S to X is a function from S to the space of all
subsets of X. The upper inverse Ψu of a subset A ⊆ X is
Ψu(A) = {s ∈ S : Ψ(s) ⊆ A}.
The lower inverse Ψl of a subset A ⊆ X is
Ψl(A) = {s ∈ S : Ψ(s) ∩A 6= ∅}.
The correspondence Ψ is
1. weakly measurable, if Ψl(O) ∈ S for each open subset O ⊆ X;
2. measurable, if Ψl(K) ∈ S for each closed subset K ⊆ X.
The graph of Ψ is denoted by Gr(Ψ) = {(s, x) ∈ S × X : s ∈ S, x ∈ Ψ(s)}. The
correspondence Ψ is said to have a measurable graph if Gr(Ψ) ∈ S ⊗ B(X).
If S is a topological space, then Ψ is
1. upper hemicontinuous, if Ψu(O) is open for each open subset O ⊆ X;
2. lower hemicontinuous, if Ψl(O) is open for each open subset O ⊆ X.
The following lemma shows that the space of nonempty compact subsets of a
Polish space is still Polish under the Hausdorff metric topology.
Lemma 1. Suppose that X is a Polish space and K is the set of all nonempty
compact subsets of X endowed with the Hausdorff metric topology. Then K is a
Polish space.
16
Proof. By Theorem 3.88 (2) of Aliprantis and Border (2006), K is complete. In
addition, Corollary 3.90 and Theorem 3.91 of Aliprantis and Border (2006) imply
that K is separable. Thus, K is a Polish space.
The following two lemmas present some basic measurability and continuity
properties for correspondences.
Lemma 2. Let (S,S) be a measurable space, X a Polish space endowed with the
Borel σ-algebra B(X), and K the space of nonempty compact subsets of X endowed
with its Hausdorff metric topology. Suppose that Ψ: S → X is a nonempty and
closed valued correspondence.
1. If Ψ is weakly measurable, then it has a measurable graph.
2. If Ψ is compact valued, then the following statements are equivalent.
(a) The correspondence Ψ is weakly measurable.
(b) The correspondence Ψ is measurable.
(c) The function f : S → K, defined by f(s) = Ψ(s), is Borel measurable.
3. Suppose that S is a topological space. If Ψ is compact valued, then the
function f : S → K defined by f(s) = Ψ(s) is continuous if and only if the
correspondence Ψ is continuous.
4. Suppose that (S,S, λ) is a complete probability space. Then Ψ is S-measurable
if and only if it has a measurable graph.
5. For a correspondence Ψ: S → X between two Polish spaces, the following
statements are equivalent.
(a) The correspondence Ψ is upper hemicontinuous at a point s ∈ S and
Ψ(s) is compact.
(b) If a sequence (sn, xn) in the graph of Ψ satisfies sn → s, then the sequence
{xn} has a limit in Ψ(s).
6. For a correspondence Ψ: S → X between two Polish spaces, the following
statements are equivalent.
(a) The correspondence Ψ is lower hemicontinuous at a point s ∈ S.
(b) If sn → s, then for each x ∈ Ψ(s), there exist a subsequence {snk} of
{sn} and elements xk ∈ Ψ(snk) for each k such that xk → x.
7. Given correspondences F : X → Y and G : Y → Z, the composition F and G
is defined by
G(F (x)) = ∪y∈F (x)G(y).
17
The composition of upper hemicontinuous correspondences is upper hemicon-
tinuous. The composition of lower hemicontinuous correspondences is lower
hemicontinuous.
Proof. Properties (1), (2), (3), (5), (6) and (7) are Theorems 18.6, 18.10, 17.15,
17.20, 17.21 and 17.23 of Aliprantis and Border (2006), respectively. Property (4)
follows from Proposition 4 of Hildenbrand (1974, p.61).
Lemma 3. 1. A correspondence Ψ from a measurable space (S,S) into a topo-
logical space X is weakly measurable if and only if its closure correspondence
Ψ is weakly measurable, where for each s ∈ S, Ψ(s) = Ψ(s) and Ψ(s) is the
closure of the set Ψ(s) in X.
2. For a sequence {Ψm} of correspondences from a measurable space (S,S)
into a Polish space, the union correspondence Ψ(s) = ∪m≥1Ψm(s) is weakly
measurable if each Ψm is weakly measurable. If each Ψm is weakly measurable
and compact valued, then the intersection correspondence Φ(s) = ∩m≥1Ψm(s)
is weakly measurable.
3. A weakly measurable, nonempty and closed valued correspondence from a
measurable space into a Polish space admits a measurable selection.
4. A correspondence with closed graph between compact metric spaces is mea-
surable.
5. A nonempty and compact valued correspondence Ψ from a measurable space
(S,S) into a Polish space is weakly measurable if and only if there exists
a sequence {ψ1, ψ2. . . .} of measurable selections of Ψ such that Ψ(s) =
{ψ1(s), ψ2(s), . . .} for each s ∈ S.
6. The image of a compact set under a compact valued upper hemicontinuous
correspondence is compact.22 If the domain is compact, then the graph of a
compact valued upper hemicontinuous correspondence is compact.
7. The intersection of a correspondence with closed graph and an upper hemi-
continuous compact valued correspondence is upper hemicontinuous.
8. If the correspondence Ψ: S → Rl is compact valued and upper hemi-
continuous, then the convex hull of Ψ is also compact valued and upper
hemicontinuous.
Proof. Properties (1)-(7) are Lemmas 18.3 and 18.4, Theorems 18.13 and 18.20,
Corollary 18.15, Lemma 17.8 and Theorem 17.25 in Aliprantis and Border (2006),
respectively. Property 8 is Proposition 6 in Hildenbrand (1974, p.26).
22Given a correspondence F : X → Y and a subset A of X, the image of A under F is defined to bethe set ∪x∈AF (x).
18
Parts 1 and 2 of the following lemma are the standard Lusin’s Theorem
and Michael’s continuous selection theorem, while the other two parts are about
properties involving mixture of measurability and continuity of correspondences.
Lemma 4. 1. Lusin’s Theorem: Suppose that S is a Borel subset of a Polish
space, λ is a Borel probability measure on S and S is the completion of B(S)
under λ. Let X be a Polish space. If f is an S-measurable mapping from S to
X, then for any ε > 0, there exists a compact subset S1 ⊆ S with λ(S\S1) < ε
such that the restriction of f to S1 is continuous.
2. Michael’s continuous selection theorem: Let S be a metrizable space, and X a
complete metrizable closed subset of some locally convex space. Suppose that
F : S → X is a lower hemicontinuous, nonempty, convex and closed valued
correspondence. Then there exists a continuous mapping f : S → X such that
f(s) ∈ F (s) for all s ∈ S.
3. Let (S,S, λ) be a finite measure space, X a Polish space, and Y a locally
convex linear topological space. Let F : S → X be a closed-valued corre-
spondence such that Gr(F ) ∈ S ⊗ B(X), and f : Gr(F ) → Y a measurable
function which is sectionally continuous on X. Then there exists a measurable
function f ′ : S × X → Y such that (1) f ′ is sectionally continuous on
X, (2) for λ-almost all s ∈ S, f ′(s, x) = f(s, x) for all x ∈ F (s) and
f ′(s,X) ⊆ cof(s, F (s)).23
4. Let (S,S) be a measurable space, and X and Y Polish spaces. Let Ψ: S×X →M(Y ) be an S ⊗ B(X)-measurable, nonempty, convex and compact valued
correspondence which is sectionally continuous on X, where the compactness
and continuity are with respect to the weak∗ topology on M(Y ). Then there
exists an S⊗B(X)-measurable selection ψ of Ψ that is sectionally continuous
on X.
Proof. Lusin’s theorem is Theorem 7.1.13 in Bogachev (2007). Michael’s contin-
uous selection theorem can be found in Michael (1966) and the last paragraph of
Bogachev (2007, p.228). Properties (3) is Theorem 2.7 in Brown and Schreiber
(1989). Property (4) follows from Theorem 1 and the main lemma of Kucia
(1998).
The following lemma presents the convexity, compactness and continuity
properties on integration of correspondences.
23For any set A in a linear topological space, coA denotes the convex hull of A.
19
Lemma 5. Let (S,S, λ) be an atomless probability space, X a Polish space, and
F a correspondence from S to Rl. Denote∫SF (s)λ(ds) =
{∫Sf(s)λ(ds) : f is an integrable selection of F on S
}.
1. If F is measurable, nonempty and closed valued, and λ-integrably bounded by
some integrable function ψ : S → R+ in the sense that for λ-almost all s ∈ S,
‖y‖ ≤ ψ(s) for any y ∈ F (s), then∫S F (s)λ(ds) is nonempty, convex and
compact, and ∫SF (s)λ(ds) =
∫S
coF (s)λ(ds).
2. If G is a measurable, nonempty and closed valued correspondence from S ×X → Rl such that (1) G(s, ·) is upper (resp. lower) hemicontinuous on X
for all s ∈ S, and (2) G is λ-integrably bounded by some integrable function
ψ : S → R+ in the sense that for λ-almost all s ∈ S, ‖y‖ ≤ ψ(s) for any x ∈ Xand y ∈ G(s, x), then
∫S G(s, x)λ(ds) is upper (resp. lower) hemicontinuous
on X.
Proof. See Theorems 2, 3 and 4, Propositions 7 and 8, and Problem 6 in
Section D.II.4 of Hildenbrand (1974).
The following result proves a measurable version of Lyapunov’s theorem, which
is taken from Mertens (2003). Let (S,S) and (X,X ) be measurable spaces. A
transition probability from S to X is a mapping f from S to the space M(X) of
probability measures on (X,X ) such that f(B|·) : s→ f(B|s) is S-measurable for
each B ∈ X .
Lemma 6. Let f(·|s) be a transition probability from a measurable space (S,S)
to another measurable space (X,X ) (X is separable).24 Let Q be a measurable,
nonempty and compact valued correspondence from S × X to Rl, which is f -
integrable in the sense that for any measurable selection q of Q, q(·, s) is f(·|s)-absolutely integrable for any s ∈ S. Let
∫Qdf be the correspondence from S to
subsets of Rl defined by
M(s) =
(∫Qdf
)(s) =
{∫Xq(s, x)f(dx|s) : q is a measurable selection of Q
}.
Denote the graph of M by J . Let J be the restriction of the product σ-algebra
S ⊗ B(Rl) to J .
Then
24A σ-algebra is said to be separable if it is generated by a countable collection of sets.
20
1. M is a measurable, nonempty and compact valued correspondence;
2. there exists a measurable, Rl-valued function g on (X × J,X ⊗ J ) such that
g(x, e, s) ∈ Q(x, s) and e =∫X g(x, e, s)f(dx|s).
The following result is Lemma 6 of Reny and Robson (2002).
Lemma 7. Suppose that H and X are compact metric spaces. Let P : H×X → Rn
be a nonempty valued and upper hemicontinuous correspondence, and the mappings
f : H → M(X) and µ : H → 4(X) be measurable. In addition, suppose that
µ(·|h) = p(h, ·) ◦ f(·|h) such that p(h, ·) is a measurable selection of P (h, ·). Then
there exists a jointly Borel measurable selection g of P such that µ(·|h) = g(h, ·) ◦f(·|h); that is, g(h, x) = p(h, x) for f(·|h)-almost all x.
Suppose that (S1,S1) is a measurable space, S2 is a Polish space endowed with
the Borel σ-algebra, and S = S1×S2 which is endowed with the product σ-algebra
S. Let D be an S-measurable subset of S such that D(s1) is compact for any
s1 ∈ S1. The σ-algebra D is the restriction of S on D. Let X be a Polish space,
and A a D-measurable, nonempty and closed valued correspondence from D to X
which is sectionally continuous on S2. The following lemma considers the property
of upper hemicontinuity for the correspondence M as defined in Lemma 6.
Lemma 8. Let f(·|s) be a transition probability from (D,D) to M(X) such that
f(A(s)|s) = 1 for any s ∈ D, which is sectionally continuous on S2. Let G be a
bounded, measurable, nonempty, convex and compact valued correspondence from
Gr(A) to Rl, which is sectionally upper hemicontinuous on S2 ×X. Let∫Gdf be
the correspondence from D to subsets of Rl defined by
M(s) =
(∫G df
)(s) =
{∫Xg(s, x)f(dx|s) : g is a measurable selection of G
}.
Then M is S-measurable, nonempty and compact valued, and sectionally upper
hemicontinuous on S2.
Proof. Define a correspondence G : S ×X → Rl as
G =
G(s, x), if (s, x) ∈ Gr(A);
{0}, otherwise.
Then M(s) =(∫
G df)
(s) =(∫G df
)(s). The measurability, nonemptiness and
compactness follows from Lemma 6. Given s1 ∈ S1 such that (1) D(s1) 6= ∅,(2) f(s1, ·) and G(s1, ·, ·) is upper hemicontinuous. The upper hemicontinuity of
M(s1, ·) follows from Lemma 2 in Simon and Zame (1990) and Lemma 4 in Reny
and Robson (2002).
21
From now onwards, whenever we work with mappings taking values in the
spaceM(X) of Borel probability measures on some separable metric space X, the
relevant continuity or convergence is assumed to be in terms of the topology of
weak convergence of measures on M(X) unless otherwise noted.
In the following lemma, we state some properties for transition correspondences.
Lemma 9. Suppose that Y and Z are Polish spaces. Let G be a measurable,
nonempty, convex and compact valued correspondence from Y to M(Z). Define a
correspondence G′ from M(Y ) to M(Z) as
G′(ν) =
{∫Yg(y)ν(dy) : g is a Borel measurable selection of G
}.25
1. The correspondence G′ is measurable, nonempty, convex and compact valued.
2. The correspondence G is upper hemicontinuous if and only if G′ is upper
hemicontinuous. In addition, if G is continuous, then G′ is continuous.
Proof. (1) is Lemma 19.29 of Aliprantis and Border (2006). By Theorem 19.30
therein, G is upper hemicontinuous if and only if G′ is upper hemicontinuous. We
need to show that G′ is lower hemicontinuous if G is lower hemicontinuous.
Let Z be endowed with a totally bounded metric, and U(Z) the space of
bounded, real-valued and uniformly continuous functions on Z endowed with the
supremum norm, which is obviously separable. Pick a countable set {fm}m≥1 ⊆U(Z) such that {fm} is dense in the unit ball of U(Z). It follows from Theorem 15.2
of Aliprantis and Border (2006) that the weak∗ topology ofM(Z) is metrizable by
the metric dz, where
dz(µ1, µ2) =∞∑m=1
1
2m
∣∣∣∣∫Zfm(z)µ1(dz)−
∫Zfm(z)µ2(dz)
∣∣∣∣for each pair of µ1, µ2 ∈M(Z).
Suppose that {νj}j≥0 is a sequence in M(Y ) such that νj → ν0 as j → ∞.
Pick an arbitrary point µ0 ∈ G′(ν0). By the definition of G′, there exists a Borel
measurable selection g of G such that µ0 =∫Y g(y)ν0(dy).
For each k ≥ 1, by Lemma 4 (Lusin’s theorem), there exists a compact subset
Dk ⊆ Y such that g is continuous on Dk and ν0(Y \ Dk) < 13k . Define a
25The integral∫Yg(y)ν(dy) defines a Borel probability measure τ on Z such that for any Borel set C
in Z, τ(C) =∫Yg(y)(C)ν(dy). The measure τ is also equal to the Gelfand integral of g with respect to
the measure ν on Y , whenM(Z) is viewed as a set in the dual space of the space of bounded continuousfunctions on Z; see Definition 11.49 of Aliprantis and Border (2006).
22
correspondence Gk : Y →M(X) as follows:
Gk(y) =
{g(y)}, y ∈ Dk;
G(y), y ∈ Y \Dk.
Then Gk is nonempty, convex and compact valued, and lower hemicontinuous.
By Theorem 3.22 in Aliprantis and Border (2006), Y is paracompact. Then by
Lemma 4 (Michael’s selection theorem), it has a continuous selection gk.
For each k, since νj → ν0 and gk is continuous,∫Y gk(y)νj(dy)→
∫Y gk(y)ν0(dy)
in the sense that for any m ≥ 1,∫Y
∫Zfm(z)gk(dz|y)νj(dy)→
∫Y
∫Zfm(z)gk(dz|y)ν0(dy).
Thus, there exists a point νjk such that {jk} is an increasing sequence and
dz
(∫Ygk(y)νjk(dy),
∫Ygk(y)ν0(dy)
)<
1
3k.
In addition, since gk coincides with g on Dk and ν0(Y \Dk) <13k ,
dz
(∫Ygk(y)ν0(dy),
∫Yg(y)ν0(dy)
)<
2
3k.
Thus,
dz
(∫Ygk(y)νjk(dy),
∫Yg(y)ν0(dy)
)<
1
k.
Let µjk =∫Y gk(y)νjk(dy) for each k. Then µjk ∈ G′(νjk) and µjk → µ0 as k →∞.
By Lemma 2, G′ is lower hemicontinuous.
The next lemma presents some properties for the composition of two transition
correspondences in terms of the product of transition probabilities.
Lemma 10. Let X, Y and Z be Polish spaces, and G a measurable, nonempty and
compact valued correspondence from X to M(Y ). Suppose that F is a measurable,
nonempty, convex and compact valued correspondence from X×Y toM(Z). Define
a correspondence Π from X to M(Y × Z) as follows:
Π(x) = {g(x) � f(x) : g is a Borel measurable selection of G,
f is a Borel measurable selection of F}.
1. If F is sectionally continuous on Y , then Π is a measurable, nonempty and
compact valued correspondence.
23
2. If there exists a function g from X toM(Y ) such that G(x) = {g(x)} for any
x ∈ X, then Π is a measurable, nonempty and compact valued correspondence.
3. If both G and F are continuous correspondences, then Π is a nonempty and
compact valued, and continuous correspondence.26
4. If G(x) ≡ {λ} for some fixed Borel probability measure λ ∈ M(Y ) and F is
sectionally continuous on X, then Π is a continuous, nonempty and compact
valued correspondence.
Proof. (1) Define three correspondences F : X×Y →M(Y ×Z), F : M(X×Y )→M(Y × Z) and F : X ×M(Y )→M(Y × Z) as follows:
F (x, y) = {δy ⊗ µ : µ ∈ F (x, y)},
F (τ) =
{∫X×Y
f(x, y)τ(d(x, y)) : f is a Borel measurable selection of F
},
F (x, µ) = F (δx ⊗ µ).
Since F is measurable, nonempty, convex and compact valued, F is measurable,
nonempty, convex and compact valued. By Lemma 9, the correspondence F is
measurable, nonempty, convex and compact valued, and F (x, ·) is continuous on
M(Y ) for any x ∈ X.
Since G is measurable and compact valued, there exists a sequence of Borel
measurable selections {gk}k≥1 of G such that G(x) = {g1(x), g2(x), . . .} for any
x ∈ X by Lemma 3 (5). For each k ≥ 1, define a correspondence Πk from X to
M(Y ×Z) by letting Πk(x) = F (x, gk(x)) = F (x⊗gk(x)). Then Πk is measurable,
nonempty, convex and compact valued.
Fix any x ∈ X. It is clear that Π(x) = F (x,G(x)) is a nonempty valued. Since
G(x) is compact, and F (x, ·) is compact valued and continuous, Π(x) is compact
by Lemma 3. Thus, the closure⋃k≥1 Πk(x) of
⋃k≥1 Πk(x) is a subset of Π(x).
Fix any x ∈ X and τ ∈ Π(x). There exists a point ν ∈ G(x) such that
τ ∈ F (x, ν). Since {gk(x)}k≥1 is dense in G(x), it has a subsequence {gkm(x)}such that gkm(x)→ ν. As F (x, ·) is continuous, F (x, gkm(x))→ F (x, ν). That is,
τ ∈⋃k≥1
F (x, gk(x)) =⋃k≥1
Πk(x).
Therefore,⋃k≥1 Πk(x) = Π(x) for any x ∈ X. Lemma 3 (1) and (2) imply that Π
is measurable.
26In Lemma 29 of Harris, Reny and Robson (1995), they showed that Π is upper hemicontinuous ifboth G and F are upper hemicontinuous.
24
(2) As in (1), the correspondence F is measurable, nonempty, convex and
compact valued. If G = {g} for some measurable function g, then Π(x) =
F (δx ⊗ g(x)), which is measurable, nonempty and compact valued.
(3) We continue to work with the two correspondences F : X×Y →M(Y ×Z)
and F : M(X×Y )→M(Y ×Z) as in Part (1). By the condition on F , it is obvious
that the correspondence F is continuous, nonempty, convex and compact valued.
Lemma 9 implies the properties for the correspondence F . Define a correspondence
G : X → M(X × Y ) as G(x) = δx ⊗ G(x). Since G and F are both nonempty
valued, Π(x) = F (G(x)) is nonempty. As G is compact valued and F is continuous,
Π is compact valued by Lemma 3. As G and F are both continuous, Π is continuous
by Lemma 2 (7).
(4) Let Y ′ = Y and define a correspondence F : X×Y →M(Y ′×Z) as follows:
F (x, y) = {δy ⊗ µ : µ ∈ F (x, y)}.
Then F is also measurable, nonempty, convex and compact valued, and sectionally
upper hemicontinuous on X.
Let d be a totally bounded metric on Y ′ × Z, and U(Y ′ × Z) the space of
bounded, real-valued and uniformly continuous functions on Y ′×Z endowed with
the supremum norm, which is obviously separable. It follows from Theorem 15.2 of
Aliprantis and Border (2006) that the space of Borel probability measures on Y ′×Zwith the topology of weak convergence of measures can be viewed as a subspace
of the dual space of U(Y ′ × Z) with the weak* topology. By Corollary 18.37 of
Aliprantis and Border (2006), Π(x) =∫Y F (x, y)λ(dy) is nonempty, convex and
compact for any x ∈ X.27
Now we shall show the upper hemicontinuity. If xn → x0 and µn ∈ Π(xn), we
need to prove that there exists some µ0 ∈ Π(x0) such that a subsequence of {µn}weakly converges to µ0. Suppose that for n ≥ 1, fn is a Borel measurable selection
of F (xn, ·) such that µn = λ � fn.
Fix any y ∈ Y , let J(y) = co{fn(y)⊗δxn}n≥1, which is the closure of the convex
hull of {fn(y)⊗δxn}n≥1. It is obvious that J(y) is nonempty and convex. It is also
clear that J(y) is the closure of the countable set{n∑i=1
αifi(y)⊗ δxi : n ≥ 1, αi ∈ Q+, i = 1, . . . , n,
n∑i=1
αi = 1
},
27Note that the integral∫YF (x, y)λ(dy) can be viewed as the Gelfand integral of the correspondence
in the dual space of U(Y ′ × Z); see Definition 18.36 of Aliprantis and Border (2006).
25
where Q+ is the set of non-negative rational numbers. Let F ′(x) = {µ ⊗ δx :
µ ∈ F (x, y)} for any x ∈ X. Then, F ′ is continuous on X. Since {xn : n ≥0} is a compact set, Lemma 3 (6) implies that
⋃n≥0 F
′(xn) is compact. Hence,
{fn(y)⊗ δxn}n≥1 is relatively compact. By Theorem 5.22 of Aliprantis and Border
(2006), {fn(y)⊗δxn}n≥1 is tight. That is, for any positive real number ε, there is a
compact set Kε in Z ×X such that for any n ≥ 1, fn(y)⊗ δxn (Kε) > 1− ε. Thus,∑ni=1 αifi(y) ⊗ δxi (Kε) > 1 − ε for any n ≥ 1, and for any αi ∈ Q+, i = 1, . . . , n
with∑n
i=1 αi = 1. Hence, J(y) is compact by Theorem 5.22 of Aliprantis and
Border (2006) again.
For any n ≥ 1, and for any αi ∈ Q+, i = 1, . . . , n with∑n
i=1 αi = 1, it is clear
that∑n
i=1 αifi(y) ⊗ δxi is measurable in y ∈ Y . Lemma 3 (5) implies that J is
also a measurable correspondence from Y to M(Z ×X). By the argument in the
second paragraph of the proof of Part (4), the set
Λ = {λ � ζ : ζ is a Borel measurable selection of J}
is compact.
Since λ�(fn⊗δxn) ∈ Λ for each n, there exists some Borel measurable selection
ζ of J such that a subsequence of {λ � (fn ⊗ δxn)}, say itself, weakly converges to
λ � ζ ∈ Λ. Let ζX(y) be the marginal probability of ζ(y) on X for each y. Since xn
converges to x0, ζX(y) = δx0 for λ-almost all y ∈ Y . As a result, there exists a Borel
measurable function f0 such that ζ = f0 ⊗ δx0 , where f0(y) ∈ coLsn{fn(y)} for λ-
almost all y ∈ Y . Since F is convex and compact valued, and upper hemicontinuous
on X, f0 is a measurable selection of F (x0, ·). Let µ0 = λ � f0. Then µn weakly
converges to µ0, which implies that Π is upper hemicontinuous.
Next we shall show the lower hemicontinuity of Π. Suppose that xn → x0 and
µ0 ∈ Π(x0). Then there exists a Borel measurable selection f0 of F (x0, ·) such that
µ0 = λ�f0. Since F is sectionally lower hemicontinuous on X and compact valued,
for each n ≥ 1, there exists a measurable selection fn for F (xn, ·) such that fn(y)
weakly converges to f0(y) for each y ∈ Y .28 Denote µn = λ � fn. For any bounded
continuous function ψ on Y ×Z,∫Z ψ(y, z)fn(dz|y) converges to
∫Z ψ(y, z)f0(dz|y)
for any y ∈ Y . Thus, we have∫Y×Z
ψ(y, z)µn(d(y, z)) =
∫Y
∫Zψ(y, z)fn(dz|y)λ(dy)
→∫Y
∫Zψ(y, z)f0(dz|y)λ(dy)
by the Dominated Convergence Theorem. Therefore, Π is lower hemicontinuous.
28See Proposition 4.2 in Sun (1997). Notice that the atomless Loeb probability measurable spaceassumption is not needed for the result of lower hemicontinuity as in Theorem 10 therein.
26
The proof is thus complete.
The following result presents a variant of Lemma 6 in terms of transition
correspondences.
Lemma 11. Let X and Y be Polish spaces, and Z a compact subset of Rl+.
Let G be a measurable, nonempty and compact valued correspondence from X to
M(Y ). Suppose that F is a measurable, nonempty, convex and compact valued
correspondence from X × Y to Z. Define a correspondence Π from X to Z as
follows:
Π(x) = {∫Yf(x, y)g(dy|x) : g is a Borel measurable selection of G,
f is a Borel measurable selection of F}.
If F is sectionally continuous on Y , then
1. the correspondence F : X × M(Y ) → Z as F (x, ν) =∫Y F (x, y)ν(dy) is
sectionally continuous on M(Y ); and
2. Π is a measurable, nonempty and compact valued correspondence.
3. If F and G are both continuous, then Π is continuous.
Proof. (1) For any fixed x ∈ X, the upper hemicontinuity of F (x, ·) follows from
Lemma 8.
Next, we shall show the lower hemicontinuity. Fix any x ∈ X. Suppose that
{νj}j≥0 is a sequence in M(Y ) such that νj → ν0 as j → ∞. Pick an arbitrary
point z0 ∈ F (x, ν0). Then there exists a Borel measurable selection f of F (x, ·)such that z0 =
∫Y f(y)ν0(dy).
By Lemma 4 (Lusin’s theorem), for each k ≥ 1, there exists a compact subset
Dk ⊆ Y such that f is continuous on Dk and ν0(Y \Dk) <1
3kM , where M > 0 is
the bound of Z. Define a correspondence Fk : Y → Z as follows:
Fk(y) =
{f(y)}, y ∈ Dk;
F (x, y), y ∈ Y \Dk.
Then Fk is nonempty, convex and compact valued, and lower hemicontinuous.
By Theorem 3.22 in Aliprantis and Border (2006), Y is paracompact. Then by
Lemma 4 (Michael’s selection theorem), Fk has a continuous selection fk.
For each k, since νj → ν0 and fk is bounded and continuous,∫Y fk(y)νj(dy)→∫
Y fk(y)ν0(dy) as j → ∞. Thus, there exists a subsequence {νjk} such that {jk}
27
is an increasing sequence, and for each k ≥ 1,∥∥∥∥∫Yfk(y)νjk(dy)−
∫Yfk(y)ν0(dy)
∥∥∥∥ < 1
3k,
where ‖ · ‖ is the Euclidean norm on Rl.Since fk coincides with f on Dk, ν0(Y \Dk) <
13kM , and Z is bounded by M ,∥∥∥∥∫
Yfk(y)ν0(dy)−
∫Yf(y)ν0(dy)
∥∥∥∥ < 2
3k.
Thus, ∥∥∥∥∫Yfk(y)νjk(dy)−
∫Yf(y)ν0(dy)
∥∥∥∥ < 1
k.
Let zjk =∫Y fk(y)νjk(dy) for each k. Then zjk ∈ F (x, νjk) and zjk → z0 as k →∞.
By Lemma 2, F (x, ·) is lower hemicontinuous.
(2) Since G is measurable and compact valued, there exists a sequence of Borel
measurable selections {gk}k≥1 of G such that G(x) = {g1(x), g2(x), . . .} for any
x ∈ X by Lemma 3 (5). For each k ≥ 1, define a correspondence Πk from X to Z
by letting Πk(x) = F (x, gk(x)) =∫Y F (x, y)gk(dy|x). Since F is convex valued, so
is Πk. By Lemma 6, Πk is also measurable, nonempty and compact valued.
Fix any x ∈ X. It is clear that Π(x) = F (x,G(x)) is nonempty valued. Since
G(x) is compact, and F (x, ·) is compact valued and continuous, Π(x) is compact
by Lemma 3. Thus,⋃k≥1 Πk(x) ⊆ Π(x).
Fix any x ∈ X and z ∈ Π(x). There exists a point ν ∈ G(x) such that
z ∈ F (x, ν). Since {gk(x)}k≥1 is dense in G(x), it has a subsequence {gkm(x)}such that gkm(x)→ ν. As F (x, ·) is continuous, F (x, gkm(x))→ F (x, ν). That is,
z ∈⋃k≥1
F (x, gk(x)) =⋃k≥1
Πk(x).
Therefore,⋃k≥1 Πk(x) = Π(x) for any x ∈ X. Lemma 3 (1) and (2) imply that Π
is measurable.
(3) Define a correspondence F : M(X × Y )→ Z as follows:
F (τ) =
{∫X×Y
f(x, y)τ(d(x, y)) : f is a Borel measurable selection of F
}.
By (1), F is continuous. Define a correspondence G : X →M(X × Y ) as G(x) =
{δx⊗ ν : ν ∈ G(x)}. Since G and F are both nonempty valued, Π(x) = F (G(x)) is
28
nonempty. As G is compact valued and F is continuous, Π is compact valued by
Lemma 3. As G and F are both continuous, Π is continuous by Lemma 2 (7).
The following lemma shows that a measurable and sectionally continuous
correspondence on a product space is approximately continuous on the product
space.
Lemma 12. Let S, X and Y be Polish spaces endowed with the Borel σ-algebras,
and λ a Borel probability measure on S. Denote S as the completion of the Borel
σ-algebra B(S) of S under the probability measure λ. Suppose that D is a B(S)⊗B(Y )-measurable subset of S × Y , where D(s) is nonempty and compact for all
s ∈ S. Let A be a nonempty and compact valued correspondence from D to X,
which is sectionally continuous on Y and has a B(S × Y ×X)-measurable graph.
Then
(i) A(s) = Gr(A(s, ·)) is an S-measurable mapping from S to the set of nonempty
and compact subsets KY×X of Y ×X;
(ii) there exist countably many disjoint compact subsets {Sm}m≥1 of S such that
(1) λ(∪m≥1Sm) = 1, and (2) for each m ≥ 1, Dm = D∩(Sm×Y ) is compact,
and A is nonempty and compact valued, and continuous on each Dm.
Proof. (i)A(s, ·) is continuous andD(s) is compact, Gr(A(s, ·)) ⊆ Y×X is compact
by Lemma 3. Thus, A is nonempty and compact valued. Since A has a measurable
graph, A is an S-measurable mapping from S to the set of nonempty and compact
subsets KY×X of Y ×X by Lemma 2 (4).
(ii) Define a correspondence D from S to Y such that D(s) = {y ∈ Y : (s, y) ∈D}. Then D is nonempty and compact valued. As in (i), D is S-measurable.
By Lemma 4 (Lusin’s Theorem), there exists a compact subset S1 ⊆ S such that
λ(S \ S1) < 12 , D and A are continuous functions on S1. By Lemma 2 (3), D and
A are continuous correspondences on S1. Let D1 = {(s, y) ∈ D : s ∈ S1, y ∈ D(s)}.Since S1 is compact and D is continuous, D1 is compact (see Lemma 3 (6)).
Following the same procedure, for any m ≥ 1, there exists a compact subset
Sm ⊆ S such that (1) Sm∩(∪1≤k≤m−1Sk) = ∅ and Dm = D∩(Sm×Y ) is compact,
(2) λ(Sm) > 0 and λ (S \ (∪1≤k≤mSm)) < 12m , and (3) A is nonempty and compact
valued, and continuous on Dm. This completes the proof.
The lemma below states an equivalence property for the weak convergence of
Borel probability measures obtained from the product of transition probabilities.
Lemma 13. Let S and X be Polish spaces, and λ a Borel probability measure on
S. Suppose that {Sk}k≥1 is a sequence of disjoint compact subsets of S such that
λ(∪k≥1Sk) = 1. For each k, define a probability measure on Sk as λk(D) = λ(D)λ(Sk)
29
for any measurable subset D ⊆ Sk. Let {νm}m≥0 be a sequence of transition
probabilities from S to M(X), and τm = λ � νm for any m ≥ 0. Then τm weakly
converges to τ0 if and only if λk � νm weakly converges to λk � ν0 for each k ≥ 1.
Proof. First, we assume that τm weakly converges to τ0. For any closed subset
E ⊆ Sk×X, we have lim supm→∞ τm(E) ≤ τ0(E). That is, lim supm→∞ λ�νm(E) ≤λ � ν0(E). For any k, 1
λ(Sk) lim supm→∞ λ � νm(E) ≤ 1λ(Sk)λ � ν0(E), which implies
that lim supm→∞ λk � νm(E) ≤ λk � ν0(E). Thus, λk � νm weakly converges to
λk � ν0 for each k ≥ 1.
Second, we consider the case that λk � νm weakly converges to λk � ν0 for each
k ≥ 1. For any closed subset E ⊆ S×X, let Ek = E∩(Sk×X) for each k ≥ 1. Then
{Ek} are disjoint closed subsets and lim supm→∞ λk � νm(Ek) ≤ λk � ν0(Ek). Since
λk � νm(E′) = 1λ(Sk)λ � νm(E′) for any k, m and measurable subset E′ ⊆ Sk ×X,
we have that lim supm→∞ λ � νm(Ek) ≤ λ � ν0(Ek). Thus,∑k≥1
lim supm→∞
λ � νm(Ek) ≤∑k≥1
λ � ν0(Ek) = λ � ν0(E).
Since the limit superior is subadditive, we have∑k≥1
lim supm→∞
λ � νm(Ek) ≥ lim supm→∞
∑k≥1
λ � νm(Ek) = lim supm→∞
λ � νm(E).
Therefore, lim supm→∞ λ � νm(E) ≤ λ � ν0(E), which implies that τm weakly
converges to τ0.
The following is a key lemma that allows one to drop the continuity condition
on the state variables through a reference measure in Theorem 1.
Lemma 14. Suppose that X,Y and S are Polish spaces and Z is a compact metric
space. Let λ be a Borel probability measure on S, and A a nonempty and compact
valued correspondence from Z ×S to X which is sectionally upper hemicontinuous
on Z and has a B(Z×S×X)-measurable graph. Let G be a nonempty and compact
valued, continuous correspondence from Z to M(X ×S). We assume that for any
z ∈ Z and τ ∈ G(z), the marginal of τ on S is λ and τ(Gr(A(z, ·))) = 1. Let F be a
measurable, nonempty, convex and compact valued correspondence from Gr(A)→M(Y ) such that F is sectionally continuous on Z ×X. Define a correspondence
Π from Z to M(X × S × Y ) by letting
Π(z) = {g(z) � f(z, ·) : g is a Borel measurable selection of G,
f is a Borel measurable selection of F}.
Then the correspondence Π is nonempty and compact valued, and continuous.
30
Proof. Let S be the completion of B(S) under the probability measure λ. By
Lemma 12, A(s) = Gr(A(s, ·)) can be viewed as an S-measurable mapping from
S to the set of nonempty and compact subsets KZ×X of Z × X. For any s ∈ S,
the correspondence Fs = F (·, s) is continuous on A(s). By Lemma 4, there exists
a measurable, nonempty and compact valued correspondence F from Z ×X × StoM(Y ) and a Borel measurable subset S′ of S with λ(S′) = 1 such that for each
s ∈ S′, Fs is continuous on Z ×X, and the restriction of Fs to A(s) is Fs.
By Lemma 4 (Lusin’s theorem), there exists a compact subset S1 ⊆ S′ such
that A is continuous on S1 and λ(S1) > 12 . Let K1 = A(S1). Then K1 ⊆ Z ×X is
compact.
Let C(K1,KM(Y )) be the space of continuous functions from K1 to KM(Y ),
where KM(Y ) is the set of nonempty and compact subsets of M(Y ). Suppose
that the restriction of S on S1 is S1. Let F1 be the restriction of F to K1 × S1.
Then F1 can be viewed as an S1-measurable function from S1 to C(K1,KM(Y ))
(see Theorem 4.55 in Aliprantis and Border (2006)). Again by Lemma 4 (Lusin’s
theorem), there exists a compact subset of S1, say itself, such that λ(S1) > 12
and F1 is continuous on S1. As a result, F1 is a continuous correspondence on
Gr(A) ∩ (S1 × Z ×X), so is F . Let λ1 be a probability measure on S1 such that
λ1(D) = λ(D)λ(S1) for any measurable subset D ⊆ S1.
For any z ∈ Z and τ ∈ G(z), the definition of G implies that there exists a
transition probability ν from S to X such that λ � ν = τ . Define a correspondence
G1 from Z toM(X×S) as follows: for any z ∈ Z, G1(z) is the set of all τ1 = λ1�νsuch that τ = λ � ν ∈ G(z). It can be easily checked that G1 is also a nonempty
and compact valued, and continuous correspondence. Let
Π1(z) = {τ1 � f(z, ·) : τ1 = λ1 � ν ∈ G1(z),
f is a Borel measurable selection of F}.
By Lemma 10, Π1 is nonempty and compact valued, and continuous. Furthermore,
it is easy to see that for any z, Π1(z) coincides with the set
{(λ1 � ν) � f(z, ·) : λ � ν ∈ G(z), f is a Borel measurable selection of F}.
Repeat this procedure, one can find a sequence of compact subsets {St} such
that (1) for any t ≥ 1, St ⊆ S′, St ∩ (S1 ∪ . . . St−1) = ∅ and λ(S1 ∪ . . .∪ St) ≥ tt+1 ,
(2) F is continuous on Gr(A)∩(St×Z×X), λt is a probability measure on St such
that λt(D) = λ(D)λ(St)
for any measurable subset D ⊆ St, and (3) the correspondence
Πt(z) = {(λt � ν) � f(z, ·) : λ � ν ∈ G(z),
31
f is a Borel measurable selection of F}.
is nonempty and compact valued, and continuous.
Pick a sequence {zk}, {νk} and {fk} such that (λ�νk)�fk(zk, ·) ∈ Π(zk), zk → z0
and (λ � νk) � fk(zk, ·) weakly converges to some κ. It is easy to see that (λt � νk) �fk(zk, ·) ∈ Πt(zk) for each t. As Π1 is compact valued and continuous, it has a
subsequence, say itself, such that zk converges to some z0 ∈ Z and (λ1�νk)�fk(zk, ·)weakly converges to some (λ1 �µ1) � f1(z0, ·) ∈ Π1(z0). Repeat this procedure, one
can get a sequence of {µm} and fm. Let µ(s) = µm(s) and f(z0, s, x) = fm(z0, s, x)
for any x ∈ A(z0, s) when s ∈ Sm. By Lemma 13, (λ � µ) � f(z0, ·) = κ, which
implies that Π is upper hemicontinuous.
Similarly, the compactness and lower hemicontinuity of Π follow from the
compactness and lower hemicontinuity of Πt for each t.
The next lemma presents the convergence property for the integrals of a
sequence of functions and probability measures.
Lemma 15. Let S and X be Polish spaces, and A a measurable, nonempty and
compact valued correspondence from S to X. Suppose that λ is a Borel probability
measure on S and {νn}1≤n≤∞ is a sequence of transition probabilities from S to
M(X) such that νn(A(s)|s) = 1 for each s and n. For each n ≥ 1, let τn = λ � νn.
Assume that the sequence {τn} of Borel probability measures on S ×X converges
weakly to a Borel probability measure τ∞ on S ×X. Let {gn}1≤n≤∞ be a sequence
of functions satisfying the following three properties.
1. For each n between 1 and ∞, gn : S ×X → R+ is measurable and sectionally
continuous on X.
2. For any s ∈ S and any sequence xn → x∞ in X, gn(s, xn) → g∞(s, x∞) as
n→∞.
3. The sequence {gn}1≤n≤∞ is integrably bounded in the sense that there exists
a λ-integrable function ψ : S → R+ such that for any n, s and x, gn(s, x) ≤ψ(s).
Then we have ∫S×X
gn(s, x)τn(d(s, x))→∫S×X
g∞(s, x)τ∞(d(s, x)).
Proof. By Theorem 2.1.3 in Castaing, De Fitte and Valadier (2004), for any
integrably bounded function g : S × X → R+ which is sectionally continuous on
X, we have ∫S×X
g(s, x)τn(d(s, x))→∫S×X
g(s, x)τ∞(d(s, x)). (2)
32
Let {yn}1≤n≤∞ be a sequence such that yn = 1n and y∞ = 0. Then yn → y∞.
Define a mapping g from S×X×{y1, . . . , y∞} such that g(s, x, yn) = gn(s, x). Then
g is measurable on S and continuous on X×{y1, . . . , y∞}. Define a correspondence
G from S to X × {y1, . . . , y∞} × R+ such that
G(s) = {(x, yn, c) : c ∈ g(s, x, yn), x ∈ A(s), 1 ≤ n ≤ ∞} .
For any s, A(s) × {y1, . . . , y∞} is compact and g(s, ·, ·) is continuous. By
Lemma 3 (6), G(s) is compact. By Lemma 2 (2), G can be viewed as a measurable
mapping from S to the space of nonempty compact subsets of X ×{y1, . . . , y∞}×R+. Similarly, A can be viewed as a measurable mapping from S to the space of
nonempty compact subsets of X.
Fix an arbitrary ε > 0. By Lemma 4 (Lusin’s theorem), there exists a compact
subset S1 ⊆ S such that A and G are continuous on S1 and λ(S \S1) < ε. Without
loss of generality, we can assume that λ(S \ S1) is sufficiently small such that∫S\S1
ψ(s)λ(ds) < ε6 . Thus, for any n,∫
(S\S1)×Xψ(s)τn(d(s, x)) =
∫(S\S1)
ψ(s)νn(X)λ(ds) <ε
6.
By Lemma 3 (6), the set E = {(s, x) : s ∈ S1, x ∈ A(s)} is compact. Since G
is continuous on S1, g is continuous on E × {y1, . . . , y∞}. Since E × {y1, . . . , y∞}is compact, g is uniformly continuous on E × {y1, . . . , y∞}. Thus, there exists a
positive integer N1 > 0 such that for any n ≥ N1, |gn(s, x)− g∞(s, x)| < ε3 for any
(s, x) ∈ E.
By Equation (2), there exists a positive integer N2 such that for any n ≥ N2,∣∣∣∣∫S×X
g∞(s, x)τn(d(s, x))−∫S×X
g∞(s, x)τ∞(d(s, x))
∣∣∣∣ < ε
3.
Let N0 = max{N1, N2}. For any n ≥ N0,∣∣∣∣∫S×X
gn(s, x)τn(d(s, x))−∫S×X
g∞(s, x)τ∞(d(s, x))
∣∣∣∣≤∣∣∣∣∫S×X
gn(s, x)τn(d(s, x))−∫S×X
g∞(s, x)τn(d(s, x))
∣∣∣∣+
∣∣∣∣∫S×X
g∞(s, x)τn(d(s, x))−∫S×X
g∞(s, x)τ∞(d(s, x))
∣∣∣∣≤∣∣∣∣∫S1×X
gn(s, x)τn(d(s, x))−∫S1×X
g∞(s, x)τn(d(s, x))
∣∣∣∣+
∣∣∣∣∣∫
(S\S1)×Xgn(s, x)τn(d(s, x))−
∫(S\S1)×X
g∞(s, x)τn(d(s, x))
∣∣∣∣∣33
+
∣∣∣∣∫S×X
g∞(s, x)τn(d(s, x))−∫S×X
g∞(s, x)τ∞(d(s, x))
∣∣∣∣≤∫E|gn(s, x)− g∞(s, x)| τn(d(s, x)) + 2 ·
∫(S\S1)×X
ψ(s)τn(d(s, x))
+
∣∣∣∣∫S×X
g∞(s, x)τn(d(s, x))−∫S×X
g∞(s, x)τ∞(d(s, x))
∣∣∣∣<ε
3+ 2 · ε
6+ε
3
= ε.
This completes the proof.
5.2 Discontinuous games with endogenous stochastic
sharing rules
Simon and Zame (1990) proved the existence of a Nash equilibrium in discontinuous
games with endogenous sharing rules. In particular, they considered a static
game with a payoff correspondence P that is bounded, nonempty, convex and
compact valued, and upper hemicontinuous. They showed that there exists a
Borel measurable selection p of the payoff correspondence, namely the endogenous
sharing rule, and a mixed strategy profile α such that α is a Nash equilibrium when
players take p as the payoff function.
In this subsection, we shall consider discontinuous games with endogenous
stochastic sharing rules. That is, we allow the payoff correspondence to depend on
some state variable in a measurable way as follows:
1. let S be a Borel subset of a Polish space, Y a Polish space, and λ a Borel
probability measure on S;
2. D is a B(S)⊗ B(Y )-measurable subset of S × Y , where D(s) is compact for
all s ∈ S and λ ({s ∈ S : D(s) 6= ∅}) > 0;
3. X =∏
1≤i≤nXi, where each Xi is a Polish space;
4. for each i, Ai is a measurable, nonempty and compact valued correspondence
from D to Xi, which is sectionally continuous on Y ;
5. A =∏
1≤i≤nAi and E = Gr(A);
6. P is a bounded, measurable, nonempty, convex and compact valued corre-
spondence from E to Rn which is essentially sectionally upper hemicontinuous
on Y ×X.
A stochastic sharing rule at (s, y) ∈ D is a Borel measurable selection of the
correspondence P (s, y, ·); i.e., a Borel measurable function p : A(s, y) → Rn such
34
that p(x) ∈ P (s, y, x) for all x ∈ A(s, y). Given (s, y) ∈ D, P (s, y, ·) represents the
set of all possible payoff profiles, and a sharing rule p is a particular choice of the
payoff profile.
Now we shall prove the following proposition.
Proposition 4. There exists a B(D)-measurable, nonempty and compact valued
correspondence Φ from D to Rn × M(X) × 4(X) such that Φ is essentially
sectionally upper hemicontinuous on Y , and for λ-almost all s ∈ S with D(s) 6= ∅and y ∈ D(s), Φ(s, y) is the set of points (v, α, µ) that
1. v =∫X p(s, y, x)α(dx) such that p(s, y, ·) is a Borel measurable selection of
P (s, y, ·);29
2. α ∈ ⊗i∈IM(Ai(s, y)) is a Nash equilibrium in the subgame (s, y) with payoff
p(s, y, ·) and action space Ai(s, y) for each player i;
3. µ = p(s, y, ·) ◦ α.30
In addition, denote the restriction of Φ on the first component Rn as Φ|Rn, which
is a correspondence from D to Rn. Then Φ|Rn is bounded, measurable, nonempty
and compact valued, and essentially sectionally upper hemicontinuous on Y .
If D is a closed subset, P is upper hemicontinuous on E and Ai is continuous
on D for each i ∈ I, then Proposition 4 is reduced to be the following lemma (see
Simon and Zame (1990) and Reny and Robson (2002, Lemma 4)).
Lemma 16. Assume that D is a closed subset, P is upper hemicontinuous on E
and Ai is continuous on D for each i ∈ I. Consider the correspondence Φ: D →Rn ×M(X)×4(X) defined as follows: (v, α, µ) ∈ Φ(s, y) if
1. v =∫X p(s, y, x)α(dx) such that p(s, y, ·) is a Borel measurable selection of
P (s, y, ·);
2. α ∈ ⊗i∈IM(Ai(s, y)) is a Nash equilibrium in the subgame (s, y) with payoff
p(s, y, ·) and action space Ai(s, y) for each player i;
3. µ = p(s, y, ·) ◦ α.
Then Φ is nonempty and compact valued, and upper hemicontinuous on D.
We shall now prove Proposition 4.
Proof of Proposition 4. There exists a Borel subset S ⊆ S with λ(S) = 1 such that
D(s) 6= ∅ for each s ∈ S, and P is sectionally upper hemicontinuous on Y when it
is restricted on D ∩ (S × Y ). Without loss of generality, we assume that S = S.
29Note that we require p(s, y, ·) to be measurable for each (s, y), but p may not be jointly measurable.30The finite measure µ = p(s, y, ·) ◦ α if µ(B) =
∫Bp(s, y, x)α(dx) for any Borel subset B ⊆ X.
35
Suppose that S is the completion of B(S) under the probability measure λ.
Let D and E be the restrictions of S ⊗ B(Y ) and S ⊗ B(Y )⊗ B(X) on D and E,
respectively.
Define a correspondence D from S to Y such that D(s) = {y ∈ Y : (s, y) ∈ D}.Then D is nonempty and compact valued. By Lemma 2 (4), D is S-measurable.
Since D(s) is compact and A(s, ·) is upper hemicontinuous for any s ∈ S, E(s)
is compact by Lemma 3 (6). Define a correspondence Γ from S to Y × X × Rn
as Γ(s) = Gr(P (s, ·, ·)). For all s, P (s, ·, ·) is bounded, upper hemicontinuous and
compact valued on E(s), hence it has a compact graph. As a result, Γ is compact
valued. By Lemma 2 (1), P has an S ⊗ B(Y ×X × Rn)-measurable graph. Since
Gr(Γ) = Gr(P ), Gr(Γ) is S⊗B(Y ×X×Rn)-measurable. Due to Lemma 2 (4), the
correspondence Γ is S-measurable. We can view Γ as a function from S into the
space K of nonempty compact subsets of Y ×X ×Rn. By Lemma 1, K is a Polish
space endowed with the Hausdorff metric topology. Then by Lemma 2 (2), Γ is
an S-measurable function from S to K. One can also define a correspondence Ai
from S to Y ×X as Ai(s) = Gr(Ai(s, ·)). It is easy to show that Ai can be viewed
as an S-measurable function from S to the space of nonempty compact subsets
of Y × X, which is endowed with the Hausdorff metric topology. By a similar
argument, D can be viewed as an S-measurable function from S to the space of
nonempty compact subsets of Y .
By Lemma 4 (Lusin’s Theorem), there exists a compact subset S1 ⊆ S such
that λ(S \ S1) < 12 , Γ, D and {Ai}1≤i≤n are continuous functions on S1. By
Lemma 2 (3), Γ, D and Ai are continuous correspondences on S1. Let D1 =
{(s, y) ∈ D : s ∈ S1, y ∈ D(s)}. Since S1 is compact and D is continuous, D1 is
compact (see Lemma 3 (6)). Similarly, E1 = E∩(S1×Y ×X) is also compact. Thus,
P is an upper hemicontinuous correspondence on E1. Define a correspondence Φ1
from D1 to Rn×M(X)×4(X) as in Lemma 16, then it is nonempty and compact
valued, and upper hemicontinuous on D1.
Following the same procedure, for any m ≥ 1, there exists a compact subset
Sm ⊆ S such that (1) Sm ∩ (∪1≤k≤m−1Sk) = ∅ and Dm = D ∩ (Sm × Y ) is
compact, (2) λ(Sm) > 0 and λ (S \ (∪1≤k≤mSm)) < 12m , and (3) there is a
nonempty and compact valued, upper hemicontinuous correspondence Φm from
Dm to Rn×M(X)×4(X), which satisfies conditions (1)-(3) in Lemma 16. Thus,
we have countably many disjoint sets {Sm}m≥1 such that (1) λ(∪m≥1Sm) = 1,
(2) Φm is nonempty and compact valued, and upper hemicontinuous on each Dm,
m ≥ 1.
Since Ai is a B(S) ⊗ B(Y )-measurable, nonempty and compact valued corre-
spondence, it has a Borel measurable selection ai by Lemma 3 (3). Fix a Borel
measurable selection p of P (such a selection exists also due to Lemma 3 (3)).
36
Define a mapping (v0, α0, µ0) from D to Rn×M(X)×4(X) such that (1) αi(s, y) =
δai(s,y) and α0(s, y) = ⊗i∈Iαi(s, y); (2) v0(s, y) = p(s, y, a1(s, y). . . . , an(s, y))
and (3) µ0(s, y) = p(s, y, ·) ◦ α0. Let D0 = D \ (∪m≥1Dm) and Φ0(s, y) =
{(v0(s, y), α0(s, y), µ0(s, y))} for (s, y) ∈ D0. Then, Φ0 is B(S)⊗B(Y )-measurable,
nonempty and compact valued.
Let Φ(s, y) = Φm(s, y) if (s, y) ∈ Dm for some m ≥ 0. Then, Φ(s, y) satisfies
conditions (1)-(3) if (s, y) ∈ Dm for m ≥ 1. That is, Φ is B(D)-measurable,
nonempty and compact valued, and essentially sectionally upper hemicontinuous
on Y , and satisfies conditions (1)-(3) for λ-almost all s ∈ S.
Then consider Φ|Rn , which is the restriction of Φ on the first component Rn.
Let Φm|Rn be the restriction of Φm on the first component Rn with the domain Dm
for each m ≥ 0. It is obvious that Φ0|Rn is measurable, nonempty and compact
valued. For each m ≥ 1, Dm is compact and Φm is upper hemicontinuous and
compact valued. By Lemma 3 (6), Gr(Φm) is compact. Thus, Gr(Φm|Rn) is also
compact. By Lemma 3 (4), Φm|Rn is measurable. In addition, Φm|Rn is nonempty
and compact valued, and upper hemicontinuous on Dm. Notice that Φ|Rn(s, y) =
Φm|Rn(s, y) if (s, y) ∈ Dm for some m ≥ 0. Thus, Φ|Rn is measurable, nonempty
and compact valued, and essentially sectionally upper hemicontinuous on Y .
The proof is complete.
5.3 Proofs of Theorem 1 and Proposition 1
5.3.1 Backward induction
For any t ≥ 1, suppose that the correspondence Qt+1 from Ht to Rn is bounded,
measurable, nonempty and compact valued, and essentially sectionally upper
hemicontinuous on Xt. For any ht−1 ∈ Ht−1 and xt ∈ At(ht−1), let
Pt(ht−1, xt) =
∫St
Qt+1(ht−1, xt, st)ft0(dst|ht−1)
=
∫St
Qt+1(ht−1, xt, st)ϕt0(ht−1, st)λt(dst).
It is obvious that the correspondence Pt is measurable and nonempty valued. Since
Qt+1 is bounded, Pt is bounded. For λt-almost all st ∈ St, Qt+1(·, st) is bounded
and upper hemicontinuous on Ht(st), and ϕt0(st, ·) is continuous on Gr(At0)(st). As
ϕt0 is integrably bounded, Pt(st−1, ·) is also upper hemicontinuous on Gr(At)(st−1)
for λt−1-almost all st−1 ∈ St−1 (see Lemma 5); that is, the correspondence Pt is
essentially sectionally upper hemicontinuous on Xt. Again by Lemma 5, Pt is
convex and compact valued since λt is an atomless probability measure. That is,
Pt : Gr(At)→ Rn is a bounded, measurable, nonempty, convex and compact valued
37
correspondence which is essentially sectionally upper hemicontinuous on Xt.
By Proposition 4, there exists a bounded, measurable, nonempty and compact
valued correspondence Φt from Ht−1 to Rn × M(Xt) × 4(Xt) such that Φt is
essentially sectionally upper hemicontinuous on Xt−1, and for λt−1-almost all
ht−1 ∈ Ht−1, (v, α, µ) ∈ Φt(ht−1) if
1. v =∫At(ht−1) pt(ht−1, x)α(dx) such that pt(ht−1, ·) is a Borel measurable
selection of Pt(ht−1, ·);
2. α ∈ ⊗i∈IM(Ati(ht−1)) is a Nash equilibrium in the subgame ht−1 with payoff
pt(ht−1, ·) and action space∏i∈I Ati(ht−1);
3. µ = pt(ht−1, ·) ◦ α.
Denote the restriction of Φt on the first component Rn as Φ(Qt+1), which is
a correspondence from Ht−1 to Rn. By Proposition 4, Φ(Qt+1) is bounded,
measurable, nonempty and compact valued, and essentially sectionally upper
hemicontinuous on Xt−1.
5.3.2 Forward induction
The following proposition presents the result on the step of forward induction.
Proposition 5. For any t ≥ 1 and any Borel measurable selection qt of Φ(Qt+1),
there exists a Borel measurable selection qt+1 of Qt+1 and a Borel measurable
mapping ft : Ht−1 → ⊗i∈IM(Xti) such that for λt−1-almost all ht−1 ∈ Ht−1,
1. ft(ht−1) ∈ ⊗i∈IM(Ati(ht−1));
2. qt(ht−1) =∫At(ht−1)
∫Stqt+1(ht−1, xt, st)ft0(dst|ht−1)ft(dxt|ht−1);
3. ft(·|ht−1) is a Nash equilibrium in the subgame ht−1 with action spaces
Ati(ht−1), i ∈ I and the payoff functions∫St
qt+1(ht−1, ·, st)ft0(dst|ht−1).
Proof. We divide the proof into three steps. In step 1, we show that there exist
Borel measurable mappings ft : Ht−1 → ⊗i∈IM(Xti) and µt : Ht−1 →4(Xt) such
that (qt, ft, µt) is a selection of Φt. In step 2, we obtain a Borel measurable selection
gt of Pt such that for λt−1-almost all ht−1 ∈ Ht−1,
1. qt(ht−1) =∫At(ht−1) gt(ht−1, x)ft(dx|ht−1);
2. ft(ht−1) is a Nash equilibrium in the subgame ht−1 with payoff gt(ht−1, ·) and
action space At(ht−1);
38
In step 3, we show that there exists a Borel measurable selection qt+1 of Qt+1 such
that for all ht−1 ∈ Ht−1 and xt ∈ At(ht−1),
gt(ht−1, xt) =
∫St
qt+1(ht−1, xt, st)ft0(dst|ht−1).
Combining Steps 1-3, the proof is complete.
Step 1. Let Ψt : Gr(Φt(Qt+1))→M(Xt)×4(Xt) be
Ψt(ht−1, v) = {(α, µ) : (v, α, µ) ∈ Φt(ht−1)}.
Recall the construction of Φt and the proof of Proposition 4, Ht−1 can be divided
into countably many Borel subsets {Hmt−1}m≥0 such that
1. Ht−1 = ∪m≥0Hmt−1 and
λt−1(∪m≥1projSt−1 (Hmt−1))
λt−1(projSt−1 (Ht−1))= 1, where projSt−1(Hm
t−1)
and projSt−1(Ht−1) are projections of Hmt−1 and Ht−1 on St−1;
2. for m ≥ 1, Hmt−1 is compact, Φt is upper hemicontinuous on Hm
t−1, and Pt is
upper hemicontinuous on
{(ht−1, xt) : ht−1 ∈ Hmt−1, xt ∈ At(ht−1)};
3. there exists a Borel measurable mapping (v0, α0, µ0) from H0t−1 to Rn ×
M(Xt) × 4(Xt) such that Φt(ht−1) ≡ {(v0(ht−1), α0(ht−1), µ0(ht−1))} for
any ht−1 ∈ H0t−1.
Denote the restriction of Φt on Hmt−1 as Φm
t . For m ≥ 1, Gr(Φmt ) is compact,
and hence the correspondence Ψmt (ht−1, v) = {(α, µ) : (v, α, µ) ∈ Φm
t (ht−1)} has
a compact graph. For m ≥ 1, Ψmt is measurable by Lemma 3 (4), and has a
Borel measurable selection ψmt due to Lemma 3 (3). Define ψ0t (ht−1, v0(ht−1)) =
(α0(ht−1), µ0(ht−1)) for ht−1 ∈ H0t−1. For (ht−1, v) ∈ Gr(Φ(Qt+1)), let ψt(ht−1, v) =
ψmt (ht−1, v) if ht−1 ∈ Hmt−1. Then ψt is a Borel measurable selection of Ψt.
Given a Borel measurable selection qt of Φ(Qt+1), let
φt(ht−1) = (qt(ht−1), ψt(ht−1, qt(ht−1))).
Then φt is a Borel measurable selection of Φt. Denote Ht−1 = ∪m≥1Hmt−1.
By the construction of Φt, there exists Borel measurable mappings ft : Ht−1 →⊗i∈IM(Xti) and µt : Ht−1 →4(Xt) such that for all ht−1 ∈ Ht−1,
1. qt(ht−1) =∫At(ht−1) pt(ht−1, x)ft(dx|ht−1) such that pt(ht−1, ·) is a Borel
measurable selection of Pt(ht−1, ·);
39
2. ft(ht−1) ∈ ⊗i∈IM(Ati(ht−1)) is a Nash equilibrium in the subgame ht−1 with
payoff pt(ht−1, ·) and action space∏i∈I Ati(ht−1);
3. µt(·|ht−1) = pt(ht−1, ·) ◦ ft(·|ht−1).
Step 2. Since Pt is upper hemicontinuous on {(ht−1, xt) : ht−1 ∈ Hmt−1, xt ∈
At(ht−1)}, due to Lemma 7, there exists a Borel measurable mapping gm such
that (1) gm(ht−1, xt) ∈ Pt(ht−1, xt) for any ht−1 ∈ Hmt−1 and xt ∈ At(ht−1), and
(2) gm(ht−1, xt) = pt(ht−1, xt) for ft(·|ht−1)-almost all xt. Fix an arbitrary Borel
measurable selection g′ of Pt. Define a Borel measurable mapping from Gr(At) to
Rn as
g(ht−1, xt) =
gm(ht−1, xt) if ht−1 ∈ Hmt−1 for m ≥ 1;
g′(ht−1, xt) otherwise.
Then g is a Borel measurable selection of Pt.
In a subgame ht−1 ∈ Ht−1, let
Bti(ht−1) = {yi ∈ Ati(ht−1) :∫At(−i)(ht−1)
gi(ht−1, yi, xt(−i))ft(−i)(dxt(−i)|ht−1) >
∫At(ht−1)
pti(ht−1, xt)ft(dxt|ht−1)}.
Since g(ht−1, xt) = pt(ht−1, xt) for ft(·|ht−1)-almost all xt,∫At(ht−1)
g(ht−1, xt)ft(dxt|ht−1) =
∫At(ht−1)
pt(ht−1, xt)ft(dxt|ht−1).
Thus, Bti is a measurable correspondence from Ht−1 to Ati(ht−1). Let Bcti(ht−1) =
Ati(ht−1) \ Bti(ht−1) for each ht−1 ∈ Ht−1. Then Bcti is a measurable and closed
valued correspondence, which has a Borel measurable graph by Lemma 2. As a
result, Bti also has a Borel measurable graph. As ft(ht−1) is a Nash equilibrium
in the subgame ht−1 ∈ Ht−1 with payoff pt(ht−1, ·), fti(Bti(ht−1)|ht−1) = 0.
Denote βi(ht−1, xt) = minPti(ht−1, xt), where Pti(ht−1, xt) is the projection
of Pt(ht−1, xt) on the i-th dimension. Then the correspondence Pti is mea-
surable and compact valued, and βi is Borel measurable. Let Λi(ht−1, xt) =
{βi(ht−1, xt)} × [0, γ]n−1, where γ > 0 is the upper bound of Pt. Denote
Λ′i(ht−1, xt) = Λi(ht−1, xt) ∩ Pt(ht−1, xt). Then Λ′i is a measurable and compact
valued correspondence, and hence has a Borel measurable selection β′i. Note that
β′i is a Borel measurable selection of Pt. Let
gt(ht−1, xt) =
40
β′i(ht−1, xt) if ht−1 ∈ Ht−1, xti ∈ Bti(ht−1) and xtj /∈ Btj(ht−1), ∀j 6= i;
g(ht−1, xt) otherwise.
Notice that
{(ht−1, xt) ∈ Gr(At) : ht−1 ∈ Ht−1, xti ∈ Bti(ht−1) and xtj /∈ Btj(ht−1),∀j 6= i; }
= Gr(At) ∩ ∪i∈I
(Gr(Bti)×∏j 6=i
Xtj) \ (∪j 6=i(Gr(Btj)×∏k 6=j
Xtk))
,
which is a Borel set. As a result, gt is a Borel measurable selection of Pt. Moreover,
gt(ht−1, xt) = pt(ht−1, xt) for all ht−1 ∈ Ht−1 and ft(·|ht−1)-almost all xt.
Fix a subgame ht−1 ∈ Ht−1. We will show that ft(·|ht−1) is a Nash equilibrium
given the payoff gt(ht−1, ·) in the subgame ht−1. Suppose that player i deviates to
some action xti.
If xti ∈ Bti(ht−1), then player i’s expected payoff is∫At(−i)(ht−1)
gti(ht−1, xti, xt(−i))ft(−i)(dxt(−i)|ht−1)
=
∫∏
j 6=iBctj(ht−1)
gti(ht−1, xti, xt(−i))ft(−i)(dxt(−i)|ht−1)
=
∫∏
j 6=iBctj(ht−1)
βi(ht−1, xti, xt(−i))ft(−i)(dxt(−i)|ht−1)
≤∫∏
j 6=iBctj(ht−1)
pti(ht−1, xti, xt(−i))ft(−i)(dxt(−i)|ht−1)
=
∫At(−i)(ht−1)
pti(ht−1, xti, xt(−i))ft(−i)(dxt(−i)|ht−1)
≤∫At(ht−1)
pti(ht−1, xt)ft(dxt|ht−1)
=
∫At(ht−1)
gti(ht−1, xt)ft(dxt|ht−1).
The first and the third equalities hold since ftj(Btj(ht−1)|ht−1) = 0 for each j,
and hence ft(−i)(∏j 6=iB
ctj(ht−1)|ht−1) = ft(−i)(At(−i)(ht−1)|ht−1). The second
equality and the first inequality are due to the fact that gti(ht−1, xti, xt(−i)) =
βi(ht−1, xti, xt(−i)) = minPti(ht−1, xti, xt(−i)) ≤ pti(ht−1, xti, xt(−i)) for xt(−i) ∈∏j 6=iB
ctj(ht−1). The second inequality holds since ft(·|ht−1) is a Nash equilibrium
given the payoff pt(ht−1, ·) in the subgame ht−1. The fourth equality follows from
the fact that gt(ht−1, xt) = pt(ht−1, xt) for ft(·|ht−1)-almost all xt.
41
If xti /∈ Bti(ht−1), then player i’s expected payoff is∫At(−i)(ht−1)
gti(ht−1, xti, xt(−i))ft(−i)(dxt(−i)|ht−1)
=
∫∏
j 6=iBctj(ht−1)
gti(ht−1, xti, xt(−i))ft(−i)(dxt(−i)|ht−1)
=
∫∏
j 6=iBctj(ht−1)
gi(ht−1, xti, xt(−i))ft(−i)(dxt(−i)|ht−1)
=
∫At(−i)(ht−1)
gi(ht−1, xti, xt(−i))ft(−i)(dxt(−i)|ht−1)
≤∫At(ht−1)
pti(ht−1, xt)ft(dxt|ht−1)
=
∫At(ht−1)
gti(ht−1, xt)ft(dxt|ht−1).
The first and the third equalities hold since
ft(−i)
∏j 6=i
Bctj(ht−1)|ht−1
= ft(−i)(At(−i)(ht−1)|ht−1).
The second equality is due to the fact that gti(ht−1, xti, xt(−i)) = gi(ht−1, xti, xt(−i))
for xt(−i) ∈∏j 6=iB
ctj(ht−1). The first inequality follows from the definition of Bti,
and the fourth equality holds since gt(ht−1, xt) = pt(ht−1, xt) for ft(·|ht−1)-almost
all xt.
Thus, player i cannot improve his payoff in the subgame ht by a unilateral
change in his strategy for any i ∈ I, which implies that ft(·|ht−1) is a Nash
equilibrium given the payoff gt(ht−1, ·) in the subgame ht−1.
Step 3. For any (ht−1, xt) ∈ Gr(At),
Pt(ht−1, xt) =
∫St
Qt+1(ht−1, xt, st)ft0(dst|ht−1).
By Lemma 6, there exists a Borel measurable mapping q from Gr(Pt) × St to Rn
such that
1. q(ht−1, xt, e, st) ∈ Qt+1(ht−1, xt, st) for any (ht−1, xt, e, st) ∈ Gr(Pt)× St;
2. e =∫Stq(ht−1, xt, e, st)ft0(dst|ht−1) for any (ht−1, xt, e) ∈ Gr(Pt), where
(ht−1, xt) ∈ Gr(At).
Let
qt+1(ht−1, xt, st) = q(ht−1, xt, gt(ht−1, xt), st)
42
for any (ht−1, xt, st) ∈ Ht. Then qt+1 is a Borel measurable selection of Qt+1.
For (ht−1, xt) ∈ Gr(At),
gt(ht−1, xt) =
∫St
q(ht−1, xt, gt(ht−1, xt), st)ft0(dst|ht−1)
=
∫St
qt+1(ht−1, xt, st)ft0(dst|ht−1).
Therefore, we have a Borel measurable selection qt+1 of Qt+1, and a Borel
measurable mapping ft : Ht−1 → ⊗i∈IM(Xti) such that for all ht−1 ∈ Ht−1,
properties (1)-(3) are satisfied. The proof is complete.
If a dynamic game has only T stages for some positive integer T ≥ 1, then let
QT+1(hT ) = {u(hT )} for any hT ∈ HT , and Qt = Φ(Qt+1) for 1 ≤ t ≤ T − 1. We
can start with the backward induction from the last period and stop at the initial
period, then run the forward induction from the initial period to the last period.
Thus, the following corollary is immediate.
Corollary 4. Any finite-horizon dynamic game with the ARM condition has a
subgame-perfect equilibrium.
5.3.3 Infinite horizon case
Pick a sequence ξ = (ξ1, ξ2, . . .) such that (1) ξm is a transition probability from
Hm−1 to M(Xm) for any m ≥ 1, and (2) ξm(Am(hm−1)|hm−1) = 1 for any m ≥ 1
and hm−1 ∈ Hm−1. Denote the set of all such ξ as Υ.
Fix any t ≥ 1, define correspondences Ξtt and ∆tt as follows: in the subgame
ht−1,
Ξtt(ht−1) =M(At(ht−1))⊗ λt,
and
∆tt(ht−1) =M(At(ht−1))⊗ ft0(ht−1).
For any m1 > t, suppose that the correspondences Ξm1−1t and ∆m1−1
t have been de-
fined. Then we can define correspondences Ξm1t : Ht−1 →M
(∏t≤m≤m1
(Xm × Sm))
and ∆m1t : Ht−1 →M
(∏t≤m≤m1
(Xm × Sm))
as follows:
Ξm1t (ht−1) ={g(ht−1) � (ξm1(ht−1, ·)⊗ λm1) :
g is a Borel measurable selection of Ξm1−1t ,
ξm1 is a Borel measurable selection of M(Am1)},
43
and
∆m1t (ht−1) ={g(ht−1) � (ξm1(ht−1, ·)⊗ fm10(ht−1, ·)) :
g is a Borel measurable selection of ∆m1−1t ,
ξm1 is a Borel measurable selection of M(Am1)},
whereM(Am1) is regarded as a correspondence from Hm1−1 to the space of Borel
probability measures onXm1 . For anym1 ≥ t, let ρm1
(ht−1,ξ)∈ Ξm1
t be the probability
measure on∏t≤m≤m1
(Xm × Sm) induced by {λm}t≤m≤m1 and {ξm}t≤m≤m1 , and
%m1
(ht−1,ξ)∈ ∆m1
t be the probability measure on∏t≤m≤m1
(Xm × Sm) induced by
{fm0}t≤m≤m1 and {ξm}t≤m≤m1 . Then, Ξm1t (ht−1) is the set of all such ρm1
(ht−1,ξ),
while ∆m1t (ht−1) is the set of all such %m1
(ht−1,ξ). Note that %m1
(ht−1,ξ)∈ ∆m1
t (ht−1) if
and only if ρm1
(ht−1,ξ)∈ Ξm1
t (ht−1). Both %m1
(ht−1,ξ)and ρm1
(ht−1,ξ)can be regarded as
probability measures on Hm1(ht−1).
Similarly, let ρ(ht−1,ξ) be the probability measure on∏m≥t(Xm×Sm) induced by
{λm}m≥t and {ξm}m≥t, and %(ht−1,ξ) the probability measure on∏m≥t(Xm × Sm)
induced by {fm0}m≥t and {ξm}m≥t. Denote the correspondence
Ξt : Ht−1 →M(∏m≥t
(Xm × Sm))
as the set of all such ρ(ht−1,ξ), and
∆t : Ht−1 →M(∏m≥t
(Xm × Sm))
as the set of all such %(ht−1,ξ).
The following lemma demonstrates the relationship between %m1
(ht−1,ξ)and
ρm1
(ht−1,ξ).
Lemma 17. For any m1 ≥ t and ht−1 ∈ Ht−1,
%m1
(ht−1,ξ)=
∏t≤m≤m1
ϕm0(ht−1, ·)
◦ ρm1
(ht−1,ξ).31
31For m ≥ t ≥ 1 and ht−1 ∈ Ht−1, the function ϕm0(ht−1, ·) is defined on Hm−1(ht−1) × Sm, whichis measurable and sectionally continuous on
∏t≤k≤m−1Xk. By Lemma 4, ϕm0(ht−1, ·) can be extended
to be a measurable function ϕm0(ht−1, ·) on the product space(∏
t≤k≤m−1Xk
)×(∏
t≤k≤m Sk
), which
is also sectionally continuous on∏t≤k≤m−1Xk. Given any ξ ∈ Υ, since ρm(ht−1,ξ)
concentrates on
Hm(ht−1), ϕm0(ht−1, ·) ◦ ρm(ht−1,ξ)= ϕm0(ht−1, ·) ◦ ρm(ht−1,ξ)
. For notational simplicity, we still use
ϕm0(ht−1, ·), instead of ϕm0(ht−1, ·), to denote the above extension. Similarly, we can work with asuitable extension of the payoff function u as needed.
44
Proof. Fix ξ ∈ Υ, and Borel subsets Cm ⊆ Xm and Dm ⊆ Sm for m ≥ t. First, we
have
%t(ht−1,ξ)(Ct ×Dt) = ξt(Ct|ht−1) · ft0(Dt|ht−1)
=
∫Xt×St
1Ct×Dt(xt, st)ϕt0(ht−1, st)(ξt(ht−1)⊗ λt)(d(xt, st)),
which implies that %t(ht−1,ξ)= ϕt0(ht−1, ·) ◦ ρt(ht−1,ξ)
.32
Suppose that %m2
(ht−1,ξ)=(∏
t≤m≤m2ϕm0(ht−1, ·)
)◦ ρm2
(ht−1,ξ)for some m2 ≥ t.
Then
%m2+1(ht−1,ξ)
∏t≤m≤m2+1
(Cm ×Dm)
= %m2
(ht−1,ξ)� (ξm2+1(ht−1, ·)⊗ f(m2+1)0(ht−1, ·))
∏t≤m≤m2+1
(Cm ×Dm)
=
∫∏
t≤m≤m2(Xm×Sm)
∫Xm2+1×Sm2+1
1∏t≤m≤m2+1(Cm×Dm)(xt, . . . , xm2+1, st, . . . , sm2+1)·
ξm2+1 ⊗ f(m2+1)0(d(xm2+1, sm2+1)|ht−1, xt, . . . , xm2 , st, . . . , sm2)
%m2
(ht−1,ξ)(d(xt, . . . , xm2 , st, . . . , sm2)|ht−1)
=
∫∏
t≤m≤m2(Xm×Sm)
∫Sm2+1
∫Xm2+1
1∏t≤m≤m2+1(Cm×Dm)(xt, . . . , xm2+1, st, . . . , sm2+1)·
ϕ(m2+1)0(ht−1, xt, . . . , xm2 , st, . . . , sm2+1)ξm2+1(dxm2+1|ht−1, xt, . . . , xm2 , st, . . . , sm2)
λ(m2+1)0(dsm2+1)∏
t≤m≤m2
ϕm0(ht−1, xt, . . . , xm−1, st, . . . , sm)
ρm2
(ht−1,ξ)(d(xt, . . . , xm2 , st, . . . , sm2)|ht−1)
=
∫∏
t≤m≤m2+1(Xm×Sm)1∏
t≤m≤m2+1(Cm×Dm)(xt, . . . , xm2+1, st, . . . , sm2+1)·∏t≤m≤m2+1
ϕm0(ht−1, xt, . . . , xm−1, st, . . . , sm)ρm2+1(ht−1,ξ)
(d(xt, . . . , xm2 , st, . . . , sm2)|ht−1),
which implies that
%m2+1(ht−1,ξ)
=
∏t≤m≤m2+1
ϕm0(ht−1, ·)
◦ ρm2+1(ht−1,ξ)
.
The proof is thus complete.
The next lemma shows that the correspondences ∆m1t and ∆t are nonempty
32For a set A in a space X, 1A is the indicator function of A, which is one on A and zero on X \A.
45
and compact valued, and sectionally continuous.
Lemma 18. 1. For any t ≥ 1, the correspondence ∆m1t is nonempty and
compact valued, and sectionally continuous on Xt−1 for any m1 ≥ t.
2. For any t ≥ 1, the correspondence ∆t is nonempty and compact valued, and
sectionally continuous on Xt−1.
Proof. (1) We first show that the correspondence Ξm1t is nonempty and compact
valued, and sectionally continuous on Xt−1 for any m1 ≥ t.Consider the case m1 = t ≥ 1, where
Ξtt(ht−1) =M(At(ht−1))⊗ λt.
Since Ati is nonempty and compact valued, and sectionally continuous on Xt−1,
Ξtt is nonempty and compact valued, and sectionally continuous on Xt−1.
Now suppose that Ξm2t is nonempty and compact valued, and sectionally
continuous on Xt−1 for some m2 ≥ t ≥ 1. Notice that
Ξm2+1t (ht−1) ={g(ht−1) � (ξm2+1(ht−1, ·)⊗ λ(m2+1)) :
g is a Borel measurable selection of Ξm2t ,
ξm2+1 is a Borel measurable selection of M(Am2+1)}.
First, we claim that Ht(s0, s1, . . . , st) is compact for any (s0, s1, . . . , st) ∈ St.We prove this claim by induction.
1. Notice that H0(s0) = X0 for any s0 ∈ S0, which is compact.
2. Suppose that Hm′(s0, s1, . . . , sm′) is compact for some 0 ≤ m′ ≤ t − 1 and
any (s0, s1, . . . , sm′) ∈ Sm′.
3. Since Am′+1(·, s0, s1, . . . , sm′) is continuous and compact valued, it has a
compact graph by Lemma 3 (6), which is Hm′+1(s0, s1, . . . , sm′+1) for any
(s0, s1, . . . , sm′+1) ∈ Sm′+1.
Thus, we prove the claim.
Define a correspondence Att from Ht−1 × St to Xt as Att(ht−1, st) = At(ht−1).
Then Att is nonempty and compact valued, sectionally continuous on Xt−1, and
has a B(Xt × St)-measurable graph. Since the graph of Att(·, s0, s1, . . . , st)
is Ht(s0, s1, . . . , st) and Ht(s0, s1, . . . , st) is compact, Att(·, s0, s1, . . . , st) has a
compact graph. For any ht−1 ∈ Ht−1 and τ ∈ Ξtt(ht−1), the marginal of τ on
St is λt and τ(Gr(Att(ht−1, ·))) = 1.
46
For any m1 > t, suppose that the correspondence
Am1−1t : Ht−1 ×
∏t≤m≤m1−1
Sm →∏
t≤m≤m1−1
Xm
has been defined such that
1. it is nonempty and compact valued, sectionally upper hemicontinuous on
Xt−1, and has a B(Xm1−1 × Sm1−1)-measurable graph;
2. for any (s0, s1, . . . .sm1−1), Am1−1t (·, s0, s1, . . . .sm1−1) has a compact graph;
3. for any ht−1 ∈ Ht−1 and τ ∈ Ξm1−1t (ht−1), the marginal of τ on∏
t≤m≤m1−1 Sm is ⊗t≤m≤m1−1λm and τ(Gr(Am1−1t (ht−1, ·))) = 1.
We define a correspondence Am1t : Ht−1 ×
∏t≤m≤m1
Sm →∏t≤m≤m1
Xm as
follows:
Am1t (ht−1, st, . . . , sm1) ={(xt, . . . , xm1) :
xm1 ∈ Am1(ht−1, xt, . . . , xm1−1, st, . . . , sm1−1),
(xt, . . . , xm1−1) ∈ Am1−1t (ht−1, st, . . . , sm1−1)}.
It is obvious that Am1t is nonempty valued. For any (s0, s1, . . . , sm1), since
Am1−1t (·, s0, s1, . . . .sm1−1) has a compact graph and Am1(·, s0, s1, . . . , sm1−1) is
continuous and compact valued, Am1t (·, s0, s1, . . . .sm1) has a compact graph by
Lemma 3 (6), which implies that Am1t is compact valued and sectionally upper
hemicontinuous on Xt−1. In addition, Gr(Am1t ) = Gr(Am1) × Sm1 , which is
B(Xm1 × Sm1)-measurable. For any ht−1 ∈ Ht−1 and τ ∈ Ξm1t (ht−1), it is obvious
that the marginal of τ on∏t≤m≤m1
Sm is ⊗t≤m≤m1λm and τ(Gr(Am1t (ht−1, ·))) =
1.
By Lemma 14, Ξm2+1t is nonempty and compact valued, and sectionally
continuous on Xt−1.
Now we show that the correspondence ∆m1t is nonempty and compact valued,
and sectionally continuous on Xt−1 for any m1 ≥ t.Given st−1 and a sequence {xk0, xk1, . . . , xkt−1} ∈ Ht−1(st−1) for 1 ≤ k ≤ ∞. Let
hkt−1 = (st−1, (xk0, xk1, . . . , x
kt−1)). It is obvious that ∆m1
t is nonempty valued, we
first show that ∆m1t is sectionally upper hemicontinuous on Xt−1. Suppose that
%m1
(hkt−1,ξk)∈ ∆m1
t (hkt−1) for 1 ≤ k <∞ and (xk0, xk1, . . . , x
kt−1)→ (x∞0 , x
∞1 , . . . , x
∞t−1),
we need to show that there exists some ξ∞ such that a subsequence of %m1
(hkt−1,ξk)
weakly converges to %m1
(h∞t−1,ξ∞) and %m1
(h∞t−1,ξ∞) ∈ ∆m1
t (h∞t−1).
Since Ξm1t is sectionally upper hemicontinuous on Xt−1, there exists some ξ∞
such that a subsequence of ρm1
(hkt−1,ξk)
, say itself, weakly converges to ρm1
(h∞t−1,ξ∞) and
47
ρm1
(h∞t−1,ξ∞) ∈ Ξm1
t (h∞t−1). Then %m1
(h∞t−1,ξ∞) ∈ ∆m1
t (h∞t−1).
For any bounded continuous function ψ on∏t≤m≤m1
(Xm × Sm), let
χk(xt, . . . , xm1 , st, . . . , sm1) =
ψ(xt, . . . , xm1 , st, . . . , sm1) ·∏
t≤m≤m1
ϕm0(hkt−1, xt, . . . , xm−1, st, . . . , sm).
Then {χk} is a sequence of functions satisfying the following three properties.
1. For each k, χk is jointly measurable and sectionally continuous on∏t≤m≤m1
Xm.
2. For any (st, . . . , sm1) and any sequence (xkt , . . . , xkm1
) → (x∞t , . . . , x∞m1
) in∏t≤m≤m1
Xm, χk(xkt , . . . , x
km1, st, . . . , sm1) → χ∞(x∞t , . . . , x
∞m1, st, . . . , sm1)
as k →∞.
3. The sequence {χk}1≤k≤∞ is integrably bounded in the sense that there exists
a function χ′ :∏t≤m≤m1
Sm → R+ such that χ′ is ⊗t≤m≤m1λm-integrable
and for any k and (xt, . . . , xm1 , st, . . . , sm1), χk(xt, . . . , xm1 , st, . . . , sm1) ≤χ′(st, . . . , sm1).
By Lemma 15, as k →∞,∫∏
t≤m≤m1(Xm×Sm)
χk(xt, . . . , xm1 , st, . . . , sm1)ρm1
(hkt−1,ξk)
(d(xt, . . . , xm1 , st, . . . , sm1))
→∫∏
t≤m≤m1(Xm×Sm)
χ∞(xt, . . . , xm1 , st, . . . , sm1)ρm1
(h∞t−1,ξ∞)(d(xt, . . . , xm1 , st, . . . , sm1)).
Then by Lemma 17,∫∏
t≤m≤m1(Xm×Sm)
ψ(xt, . . . , xm1 , st, . . . , sm1)%m1
(hkt−1,ξk)
(d(xt, . . . , xm1 , st, . . . , sm1))
→∫∏
t≤m≤m1(Xm×Sm)
ψ(xt, . . . , xm1 , st, . . . , sm1)%m1
(h∞t−1,ξ∞)(d(xt, . . . , xm1 , st, . . . , sm1)),
which implies that %m1
(hkt−1,ξk)
weakly converges to %m1
(h∞t−1,ξ∞). Therefore, ∆m1
t is
sectionally upper hemicontinuous on Xt−1. If one chooses h1t−1 = h2
t−1 = · · · =
h∞t−1, then we indeed show that ∆m1t is compact valued.
In the argument above, we indeed proved that if ρm1
(hkt−1,ξk)
weakly converges to
ρm1
(h∞t−1,ξ∞), then %m1
(hkt−1,ξk)
weakly converges to %m1
(h∞t−1,ξ∞).
The left is to show that ∆m1t is sectionally lower hemicontinuous on Xt−1.
Suppose that (xk0, xk1, . . . , x
kt−1)→ (x∞0 , x
∞1 , . . . , x
∞t−1) and %m1
(h∞t−1,ξ∞) ∈ ∆m1
t (h∞t−1),
we need to show that there exists a subsequence {(xkm0 , xkm1 , . . . , xkmt−1)} of
{(xk0, xk1, . . . , xkt−1)} and %m1
(hkmt−1,ξkm )∈ ∆m1
t (hkmt−1) for each km such that %m1
(hkmt−1,ξkm )
48
weakly converges to %m1
(h∞t−1,ξ∞).
Since %m1
(h∞t−1,ξ∞) ∈ ∆m1
t (h∞t−1), we have ρm1
(h∞t−1,ξ∞) ∈ Ξm1
t (h∞t−1). Because
Ξm1t is sectionally lower hemicontinuous on Xt−1, there exists a subsequence of
{(xk0, xk1, . . . , xkt−1)}, say itself, and ρm1
(hkt−1,ξk)∈ Ξm1
t (hkt−1) for each k such that
ρm1
(hkt−1,ξk)
weakly converges to ρm1
(h∞t−1,ξ∞). As a result, %m1
(hkt−1,ξk)
weakly converges to
%m1
(h∞t−1,ξ∞), which implies that ∆m1
t is sectionally lower hemicontinuous on Xt−1.
Therefore, ∆m1t is nonempty and compact valued, and sectionally continuous
on Xt−1 for any m1 ≥ t.
(2) We show that ∆t is nonempty and compact valued, and sectionally
continuous on Xt−1.
It is obvious that ∆t is nonempty valued, we first prove that it is compact
valued.
Given ht−1 and a sequence {τk} ⊆ ∆t(ht−1), there exists a sequence of {ξk}k≥1
such that ξk = (ξk1 , ξk2 , . . .) ∈ Υ and τk = %(ht−1,ξk) for each k.
By (1), Ξtt is compact. Then there exists a measurable mapping gt such that
(1) gt = (ξ11 , . . . , ξ
1t−1, gt, ξ
1t+1, . . .) ∈ Υ, and (2) a subsequence of {ρt
(ht−1,ξk)},
say {ρt(ht−1,ξ
k1l )}l≥1, which weakly converges to ρt(ht−1,gt)
. Note that {ξkt+1} is
a Borel measurable selection of M(At+1). By Lemma 14, there is a Borel
measurable selection gt+1 of M(At+1) such that there is a subsequence of
{ρt+1(ht−1,ξ
k1l )}l≥1, say {ρt+1
(ht−1,ξk2l )}l≥1, which weakly converges to ρt+1
(ht−1,gt+1), where
gt+1 = (ξ11 , . . . , ξ
1t−1, gt, gt+1, ξ
1t+2, . . .) ∈ Υ.
Repeat this procedure, one can construct a Borel measurable mapping g such
that ρ(ht−1,ξk11 ), ρ(ht−1,ξk22 ), ρ(ht−1,ξk33 ), . . . weakly converges to ρ(ht−1,g). That is,
ρ(ht−1,g) is a convergent point of {ρ(ht−1,ξk)}, which implies that %(ht−1,g) is a
convergent point of {%(ht−1,ξk)}.The sectional upper hemicontinuity of ∆t follows a similar argument as above.
In particular, given st−1 and a sequence {xk0, xk1, . . . , xkt−1} ⊆ Ht−1(st−1) for
k ≥ 0. Let hkt−1 = (st−1, (xk0, xk1, . . . , x
kt−1)). Suppose that (xk0, x
k1, . . . , x
kt−1) →
(x00, x
01, . . . , x
0t−1). If {τk} ⊆ ∆t(h
kt−1) for k ≥ 1 and τk → τ0, then one can show
that τ0 ∈ ∆t(h0t−1) by repeating a similar argument as in the proof above.
Finally, we consider the sectional lower hemicontinuity of ∆t. Suppose that
τ0 ∈ ∆t(h0t−1). Then there exists some ξ ∈ Υ such that τ0 = %(h0t−1,ξ)
. Denote
τm = %m(h0t−1,ξ)
∈ ∆mt (h0
t−1) for m ≥ t. As ∆mt is continuous, for each m, there
exists some ξm ∈ Υ such that d(%m(hkmt−1,ξ
m), τm) ≤ 1
m for km sufficiently large, where
d is the Prokhorov metric. Let τm = %(hkmt−1,ξ
m). Then τm weakly converges to τ0,
which implies that ∆t is sectionally lower hemicontinuous.
49
Define a correspondence Qτt : Ht−1 → Rn++ as follows:
Qτt (ht−1) ={∫∏
m≥t(Xm×Sm) u(ht−1, x, s)%(ht−1,ξ)(d(x, s)) : %(ht−1,ξ) ∈ ∆t(ht−1)}; t > τ ;
Φ(Qτt+1)(ht−1) t ≤ τ.
The lemma below presents several properties of the correspondence Qτt .
Lemma 19. For any t, τ ≥ 1, Qτt is bounded, measurable, nonempty and compact
valued, and essentially sectionally upper hemicontinuous on Xt−1.
Proof. We prove the lemma in three steps.
Step 1. Fix t > τ . We will show that Qτt is bounded, nonempty and compact
valued, and sectionally upper hemicontinuous on Xt−1.
The boundedness and nonemptiness of Qτt are obvious. We shall prove that
Qτt is sectionally upper hemicontinuous on Xt−1. Given st−1 and a sequence
{xk0, xk1, . . . , xkt−1} ⊆ Ht−1(st−1) for k ≥ 0. Let hkt−1 = (st−1, (xk0, xk1, . . . , x
kt−1)).
Suppose that ak ∈ Qτt (hkt−1) for k ≥ 1, (xk0, xk1, . . . , x
kt−1) → (x0
0, x01, . . . , x
0t−1) and
ak → a0, we need to show that a0 ∈ Qτt (h0t−1).
By the definition, there exists a sequence {ξk}k≥1 such that
ak =
∫∏
m≥t(Xm×Sm)u(hkt−1, x, s)%(hkt−1,ξ
k)(d(x, s)),
where ξk = (ξk1 , ξk2 , . . .) ∈ Υ for each k. As ∆t is compact valued and sectionally
continuous on Xt−1, there exist some %(h0t−1,ξ0) ∈ ∆t(h
0t−1) and a subsequence of
%(hkt−1,ξk), say itself, which weakly converges to %(h0t−1,ξ
0) for ξ0 = (ξ01 , ξ
02 , . . .) ∈ Υ.
We shall show that
a0 =
∫∏
m≥t(Xm×Sm)u(h0
t−1, x, s)%(h0t−1,ξ0)(d(x, s)).
For this aim, we only need to show that for any δ > 0,∣∣∣∣∣a0 −∫∏
m≥t(Xm×Sm)u(h0
t−1, x, s)%(h0t−1,ξ0)(d(x, s))
∣∣∣∣∣ < δ. (3)
Since the game is continuous at infinity, there exists a positive integer M ≥ t
such that wm < 15δ for any m > M .
For each j > M , by Lemma 4, there exists a measurable selection ξ′j ofM(Aj)
such that ξ′j is sectionally continuous on Xj−1. Let µ : HM →∏m>M (Xm ×
Sm) be the transition probability which is induced by (ξ′M+1
, ξ′M+2
, . . .) and
50
{f(M+1)0, f(M+2)0, . . .}. By Lemma 10, µ is measurable and sectionally continuous
on XM . Let
VM (ht−1, xt, . . . , xM , st, . . . , sM ) =∫∏
m>M (Xm×Sm)u(ht−1, xt, . . . , xM , st, . . . , sM , x, s) dµ(x, s|ht−1, xt, . . . , xM , st, . . . , sM ).
Then VM is bounded and measurable. In addition, VM is sectionally continuous
on XM by Lemma 15.
For any k ≥ 0, we have
∣∣ ∫∏m≥t(Xm×Sm)
u(hkt−1, x, s)%(hkt−1,ξk)(d(x, s))
−∫∏
t≤m≤M (Xm×Sm)VM (hkt−1, xt, . . . , xM , st, . . . , sM )%M
(hkt−1,ξk)
(d(xt, . . . , xM , st, . . . , sM ))∣∣
≤ wM+1
<1
5δ.
Since %(hkt−1,ξk) weakly converges to %(h0t−1,ξ
0) and %M(hkt−1,ξ
k)is the marginal of
%(hkt−1,ξk) on
∏t≤m≤M (Xm×Sm) for any k ≥ 0, the sequence %M
(hkt−1,ξk)
also weakly
converges to %M(h0t−1,ξ
0). By Lemma 15, we have
|∫∏
t≤m≤M (Xm×Sm)VM (hkt−1, xt, . . . , xM , st, . . . , sM )%M
(hkt−1,ξk)
(d(xt, . . . , xM , st, . . . , sM ))
−∫∏
t≤m≤M (Xm×Sm)VM (h0
t−1, xt, . . . , xM , st, . . . , sM )%M(h0t−1,ξ0)(d(xt, . . . , xM , st, . . . , sM ))|
<1
5δ
for k ≥ K1, where K1 is a sufficiently large positive integer. In addition, there
exists a positive integer K2 such that |ak − a0| < 15δ for k ≥ K2.
Fix k > max{K1,K2}. Combining the inequalities above, we have∣∣∣∣∣∫∏
m≥t(Xm×Sm)u(h0
t−1, x, s)%(h0t−1,ξ0)(d(x, s))− a0
∣∣∣∣∣≤∣∣ ∫∏
m≥t(Xm×Sm)u(h0
t−1, x, s)%(h0t−1,ξ0)(d(x, s))
−∫∏
t≤m≤M (Xm×Sm)VM (h0
t−1, xt, . . . , xM , st, . . . , sM )%M(h0t−1,ξ0)(d(xt, . . . , xM , st, . . . , sM ))
∣∣+∣∣ ∫∏
t≤m≤M (Xm×Sm)VM (h0
t−1, xt, . . . , xM , st, . . . , sM )%M(h0t−1,ξ0)(d(xt, . . . , xM , st, . . . , sM ))
51
−∫∏
t≤m≤M (Xm×Sm)VM (hkt−1, xt, . . . , xM , st, . . . , sM )%M
(hkt−1,ξk)
(d(xt, . . . , xM , st, . . . , sM ))∣∣
+∣∣ ∫∏
t≤m≤M (Xm×Sm)VM (hkt−1, xt, . . . , xM , st, . . . , sM )%M
(hkt−1,ξk)
(d(xt, . . . , xM , st, . . . , sM ))
−∫∏
m≥t(Xm×Sm)u(hkt−1, x, s)%(hkt−1,ξ
k)(d(x, s))∣∣
+∣∣ ∫∏
m≥t(Xm×Sm)u(hkt−1, x, s)%(hkt−1,ξ
k)(d(x, s))− a0∣∣
< δ.
Thus, we proved inequality (3), which implies that Qτt is sectionally upper
hemicontinuous on Xt−1 for t > τ .
Furthermore, to prove that Qτt is compact valued, we only need to consider
the case that {xk0, xk1, . . . , xkt−1} = {x00, x
01, . . . , x
0t−1} for any k ≥ 0, and repeat the
above proof.
Step 2. Fix t > τ , we will show that Qτt is measurable.
Fix a sequence (ξ′1, ξ′2, . . .), where ξ′j is a selection ofM(Aj) measurable in sj−1
and continuous in xj−1 for each j. For any M ≥ t, let
WMM (ht−1, xt, . . . , xM , st, . . . , sM ) ={∫
∏m>M (Xm×Sm)
u(ht−1, xt, . . . , xM , st, . . . , sM , x, s)%(ht−1,xt,...,xM ,st,...,sM ,ξ′)(d(x, s))
}.
By Lemma 10, %(ht−1,xt,...,xM ,st,...,sM ,ξ′) is measurable fromHM toM(∏
m>M (Xm × Sm)),
and sectionally continuous on XM . Thus, WMM is bounded, measurable, nonempty,
convex and compact valued. By Lemma 15, WMM is sectionally continuous on XM .
Suppose that for some t ≤ j ≤M , W jM has been defined such that it is bounded,
measurable, nonempty, convex and compact valued, and sectionally continuous on
Xj . Let
W j−1M (ht−1, xt, . . . , xj−1, st, . . . , sj−1) ={∫
Xj×Sj
wjM (ht−1, xt, . . . , xj , st, . . . , sj)%j(ht−1,xt,...,xj−1,st,...,sj−1,ξ)
(d(xj , sj)) :
%j(ht−1,xt,...,xj−1,st,...,sj−1,ξ)∈ ∆j
j(ht−1, xt, . . . , xj−1, st, . . . , sj−1),
wjM is a Borel measurable selection of W jM
}.
52
Let Sj = Sj .33 Since∫
Xj×Sj
W jM (ht−1, xt, . . . , xj , st, . . . , sj)%
j(ht−1,xt,...,xj−1,st,...,sj−1,ξ)
(d(xj , sj))
=
∫Sj
∫Xj×Sj
W jM (ht−1, xt, . . . , xj , st, . . . , sj)ρ
j(ht−1,xt,...,xj−1,st,...,sj−1,ξ)
(d(xj , sj))
· ϕj0(ht−1, xt, . . . , xj−1, st, . . . , sj)λj(dsj),
we have
W j−1M (ht−1, xt, . . . , xj−1, st, . . . , sj−1) ={∫
Sj
∫Xj×Sj
wjM (ht−1, xt, . . . , xj , st, . . . , sj)ρj(ht−1,xt,...,xj−1,st,...,sj−1,ξ)
(d(xj , sj))
· ϕj0(ht−1, xt, . . . , xj−1, st, . . . , sj)λj(dsj) :
ρj(ht−1,xt,...,xj−1,st,...,sj−1,ξ)∈ Ξjj(ht−1, xt, . . . , xj−1, st, . . . , sj−1),
wjM is a Borel measurable selection of W jM
}.
Let
W jM (ht−1, xt, . . . , xj−1, st, . . . , sj) ={∫
Xj×Sj
wjM (ht−1, xt, . . . , xj , st, . . . , sj) · ρj(ht−1,xt,...,xj−1,st,...,sj−1,ξ)(d(xj , sj)) :
ρj(ht−1,xt,...,xj−1,st,...,sj−1,ξ)∈ Ξjj(ht−1, xt, . . . , xj−1, st, . . . , sj−1),
wjM is a Borel measurable selection of W jM
}.
Since W jM (ht−1, xt, . . . , xj , st, . . . , sj) is continuous in xj and does not depend on
sj , it is continuous in (xj , sj). In addition, W jM is bounded, measurable, nonempty,
convex and compact valued. By Lemma 11, W jM is bounded, measurable,
nonempty and compact valued, and sectionally continuous on Xj−1.
It is easy to see that
W j−1M (ht−1, xt, . . . , xj−1, st, . . . , sj−1) =∫
Sj
W jM (ht−1, xt, . . . , xj−1, st, . . . , sj)ϕj0(ht−1, xt, . . . , xj−1, st, . . . , sj)λj(dsj).
By Lemma 5, it is bounded, measurable, nonempty and compact valued, and
sectionally continuous on Xj−1. By induction, one can show that W t−1M is bounded,
33We will need to use Lemma 11 below, which requires the continuity of the correspondences in termsof the integrated variables. Since W j
M is only measurable, but not continuous, in sj , we add a dummy
variable sj so that W jM is trivially continuous in such a variable.
53
measurable, nonempty and compact valued, and sectionally continuous on Xt−1.
Let W t−1 = ∪M≥tW t−1M . That is, W t−1 is the closure of ∪M≥tW t−1
M , which is
measurable due to Lemma 3.
First, W t−1 ⊆ Qτt because W t−1M ⊆ Qτt for each M ≥ t and Qτt is compact
valued. Second, fix ht−1 and q ∈ Qτt (ht−1). Then there exists a mapping ξ ∈ Υ
such that
q =
∫∏
m≥t(Xm×Sm)u(ht−1, x, s)%(ht−1,ξ)(d(x, s)).
For M ≥ t, let
VM (ht−1, xt, . . . , xM , st, . . . , sM ) =∫∏
m>M (Xm×Sm)u(ht−1, xt, . . . , xM , st, . . . , sM , x, s)%(ht−1,xt,...,xM ,st,...,sM ,ξ)(x, s)
and
qM =
∫∏
t≤m≤M (Xm×Sm)VM (ht−1, x, s)%
M(ht−1,ξ)
(d(x, s)).
Hence, qM ∈ W t−1M . Because the dynamic game is continuous at infinity, qM → q,
which implies that q ∈W t−1(ht−1) and Qτt ⊆W t−1.
Therefore, W t−1 = Qτt , and hence Qτt is measurable for t > τ .
Step 3. For t ≤ τ , we can start with Qττ+1. Repeating the backward induction
in Subsection 5.3.1, we have that Qτt is also bounded, measurable, nonempty and
compact valued, and essentially sectionally upper hemicontinuous on Xt−1.
Denote
Q∞t =
Qt−1t , if ∩τ≥1 Q
τt = ∅;
∩τ≥1Qτt , otherwise.
The following three lemmas show that Q∞t (ht−1) = Φ(Q∞t+1)(ht−1) = Et(ht−1) for
λt−1-almost all ht−1 ∈ Ht−1.34
Lemma 20. 1. The correspondence Q∞t is bounded, measurable, nonempty and
compact valued, and essentially sectionally upper hemicontinuous on Xt−1.
2. For any t ≥ 1, Q∞t (ht−1) = Φ(Q∞t+1)(ht−1) for λt−1-almost all ht−1 ∈ Ht−1.
Proof. (1) It is obvious that Q∞t is bounded. By the definition of Qτt , for λt−1-
almost all ht−1 ∈ Ht−1, Qτ1t (ht−1) ⊆ Qτ2t (ht−1) for τ1 ≥ τ2. Since Qτt is nonempty
and compact valued, Q∞t = ∩τ≥1Qτt is nonempty and compact valued for λt−1-
almost all ht−1 ∈ Ht−1. If ∩τ≥1Qτt = ∅, then Q∞t = Qt−1
t . Thus, Q∞t (ht−1) is
34The proofs for Lemmas 20 and 22 follow the standard ideas with various modifications; see, forexample, Harris (1990), Harris, Reny and Robson (1995) and Mariotti (2000).
54
nonempty and compact valued for all ht−1 ∈ Ht−1. By Lemma 3 (2), ∩τ≥1Qτt is
measurable, which implies that Q∞t is measurable.
Fix any st−1 ∈ St−1 such that Qτt (·, st−1) is upper hemicontinuous on
Ht−1(st−1) for any τ . By Lemma 3 (7), Qτt (·, st−1) has a closed graph for each τ ,
which implies that Q∞t (·, st−1) has a closed graph. Referring to Lemma 3 (7) again,
Q∞t (·, st−1) is upper hemicontinuous on Ht−1(st−1). Since Qτt is essentially upper
hemicontinuous on Xt−1 for each τ , Q∞t is essentially upper upper hemicontinuous
on Xt−1.
(2) For any τ ≥ 1 and λt−1-almost all ht−1 ∈ Ht−1, Φ(Q∞t+1)(ht−1) ⊆Φ(Qτt+1)(ht−1) ⊆ Qτt (ht−1), and hence Φ(Q∞t+1)(ht−1) ⊆ Q∞t (ht−1).
The space {1, 2, . . .∞} is a countable compact set endowed with the following
metric: d(k,m) = | 1k −1m | for any 1 ≤ k,m ≤ ∞. The sequence {Qτt+1}1≤τ≤∞
can be regarded as a correspondence Qt+1 from Ht × {1, 2, . . . ,∞} to Rn, which
is measurable, nonempty and compact valued, and essentially sectionally upper
hemicontinuous on Xt×{1, 2, . . . ,∞}. The backward induction in Subsection 5.3.1
shows that Φ(Qt+1) is measurable, nonempty and compact valued, and essentially
sectionally upper hemicontinuous on Xt × {1, 2, . . . ,∞}.Since Φ(Qt+1) is essentially sectionally upper hemicontinuous onXt×{1, 2, . . . ,∞},
there exists a measurable subset St−1 ⊆ St−1 such that λt−1(St−1) = 1, and
Φ(Qt+1)(·, ·, st−1) is upper hemicontinuous for any st−1 ∈ St−1. Fix st−1 ∈St−1. For ht−1 = (xt−1, st−1) ∈ Ht−1 and a ∈ Q∞t (ht−1), by its definition,
a ∈ Qτt (ht−1) = Φ(Qτt+1)(ht−1) for τ ≥ t. Thus, a ∈ Φ(Q∞t+1)(ht−1).
In summary, Q∞t (ht−1) = Φ(Q∞t+1)(ht−1) for λt−1-almost all ht−1 ∈ Ht−1.
Though the definition of Qτt involves correlated strategies for τ < t, the
following lemma shows that one can work with mixed strategies in terms of
equilibrium payoffs via the combination of backward and forward inductions in
multiple steps.
Lemma 21. If ct is a measurable selection of Φ(Q∞t+1), then ct(ht−1) is a subgame-
perfect equilibrium payoff vector for λt−1-almost all ht−1 ∈ Ht−1.
Proof. Without loss of generality, we only prove the case t = 1.
Suppose that c1 is a measurable selection of Φ(Q∞2 ). Apply Proposition 5
recursively to obtain Borel measurable mappings {fki}i∈I for k ≥ 1. That is,
for any k ≥ 1, there exists a Borel measurable selection ck of Q∞k such that for
λk−1-almost all hk−1 ∈ Hk−1,
1. fk(hk−1) is a Nash equilibrium in the subgame hk−1, where the action space
55
is Aki(hk−1) for player i ∈ I, and the payoff function is given by∫Sk
ck+1(hk−1, ·, sk)fk0(dsk|hk−1).
2.
ck(hk−1) =
∫Ak(hk−1)
∫Sk
ck+1(hk−1, xk, sk)fk0(dsk|hk−1)fk(dxk|hk−1).
We need to show that c1(h0) is a subgame-perfect equilibrium payoff vector for
λ0-almost all h0 ∈ H0.
Step 1. We show that for any k ≥ 1 and λk−1-almost all hk−1 ∈ Hk−1,
ck(hk−1) =
∫∏
m≥k(Xm×Sm)u(hk−1, x, s)%(hk−1,f)(d(x, s)).
Since the game is continuous at infinity, there exists some positive integer M >
k such that wM is sufficiently small. By Lemma 20, ck(hk−1) ∈ Q∞k (hk−1) =
∩τ≥1Qτk(hk−1) for λk−1-almost all hk−1 ∈ Hk−1. Since Qτk = Φτ−k+1(Qττ+1) for
k ≤ τ , ck(hk−1) ∈ ∩τ≥kΦτ−k+1(Qττ+1)(hk−1) ⊆ ΦM−k+1(QMM+1)(hk−1) for λk−1-
almost all hk−1 ∈ Hk−1. Thus, there exists a Borel measurable selection w of
QMM+1 and some ξ ∈ Υ such that for λM−1-almost all hM−1 ∈ HM−1,
i. fM (hM−1) is a Nash equilibrium in the subgame hM−1, where the action
space is AMi(hM−1) for player i ∈ I, and the payoff function is given by∫SM
w(hM−1, ·, sM )fM0(dsM |hM−1);
ii.
cM (hM−1) =
∫AM (hM−1)
∫SM
w(hM−1, xM , sM )fM0(dsM |hM−1)fM (dxM |hM−1);
iii. w(hM ) =∫∏
m≥M+1(Xm×Sm) u(hM , x, s)%(hM ,ξ)(d(x, s)).
Then for λk−1-almost all hk−1 ∈ Hk−1,
ck(hk−1) =
∫∏
m≥k(Xm×Sm)u(hk−1, x, s)%(hk−1,fM )(d(x, s)),
where fMk is fk if k ≤ M , and ξk if k ≥ M + 1. Since the game is continuous at
56
infinity, ∫∏
m≥k(Xm×Sm)u(hk−1, x, s)%(hk−1,fM )(d(x, s))
converges to ∫∏
m≥k(Xm×Sm)u(hk−1, x, s)%(hk−1,f)(d(x, s))
when M goes to infinity. Thus, for λk−1-almost all hk−1 ∈ Hk−1,
ck(ht−1) =
∫∏
m≥k(Xm×Sm)u(hk−1, x, s)%(hk−1,f)(d(x, s)). (4)
Step 2. Below, we show that {fki}i∈I is a subgame-perfect equilibrium.
Fix a player i and a strategy gi = {gki}k≥1. For each k ≥ 1, define a new
strategy fki as follows: fki = (g1i, . . . , gki, f(k+1)i, f(k+2)i, . . .). That is, we simply
replace the initial k stages of fi by gi. Denote fk = (fki , f−i).
Fix k ≥ 1 and a measurable subset Dk ⊆ Sk such that (1) and (2) of step 1
and Equation (4) hold for all sk ∈ Dk and xk ∈ Hk(sk), and λk(Dk) = 1. For each
M > k, by the Fubini property, there exists a measurable subset EMk ⊆ Sk such
that λk(EMk ) = 1 and ⊗k+1≤j≤Mλj(DM (sk)) = 1 for all sk ∈ EMk , where
DM (sk) = {(sk+1, . . . , sM ) : (sk, sk+1, . . . , sM ) ∈ DM}.
Let Dk = (∩M>kEMk ) ∩Dk. Then λk(Dk) = 1.
For any hk = (xk, sk) such that sk ∈ Dk and xk ∈ Hk(sk), we have∫
∏m≥k+1(Xm×Sm)
u(hk, x, s)%(hk,f)(d(x, s))
=
∫Ak+1(hk)
∫Sk+1
c(k+2)i(hk, xk+1, sk+1)f(k+1)0(dsk+1|hk)fk+1(dxk+1|hk)
≥∫Ak+1(hk)
∫Sk+1
c(k+2)i(hk, xk+1, sk+1)f(k+1)0(dsk+1|hk)(f(k+1)(−i) ⊗ g(k+1)i
)(dxk+1|hk)
=
∫Ak+1(hk)
∫Sk+1
∫Ak+2(hk,xk+1,sk+1)
∫Sk+2
c(k+3)i(hk, xk+1, sk+1, xk+2, sk+2)
f(k+2)0(dsk+2|hk, xk+1, sk+1)f(k+2)(−i) ⊗ f(k+2)i(dxk+2|hk, xk+1, sk+1)
f(k+1)0(dsk+1|hk)f(k+1)(−i) ⊗ g(k+1)i(dxk+1|hk)
≥∫Ak+1(hk)
∫Sk+1
∫Ak+2(hk,xk+1,sk+1)
∫Sk+2
c(k+3)i(hk, xk+1, sk+1, xk+2, sk+2)
f(k+2)0(dsk+2|hk, xk+1, sk+1)f(k+2)(−i) ⊗ g(k+2)i(dxk+2|hk, xk+1, sk+1)
f(k+1)0(dsk+1|hk)f(k+1)(−i) ⊗ g(k+1)i(dxk+1|hk)
57
=
∫∏
m≥k+1(Xm×Sm)u(hk, x, s)%(hk,fk+2)(d(x, s)).
The first and the last equalities follow from Equation (4) in the end of step 1. The
second equality is due to (2) in step 1. The first inequality is based on (1) in step 1.
The second inequality holds by the following arguments:
i. by the choice of hk and (1) in step 1, for λk+1-almost all sk+1 ∈ Sk+1 and all
xk+1 ∈ Xk+1 such that (hk, xk+1, sk+1) ∈ Hk+1, we have∫Ak+2(hk,xk+1,sk+1)
∫Sk+2
c(k+3)i(hk, xk+1, sk+1, xk+2, sk+2)
f(k+2)0(dsk+2|hk, xk+1, sk+1)f(k+2)(−i) ⊗ f(k+2)i(dxk+2|hk, xk+1, sk+1)
≥∫Ak+2(hk,xk+1,sk+1)
∫Sk+2
c(k+3)i(hk, xk+1, sk+1, xk+2, sk+2)
f(k+2)0(dsk+2|hk, xk+1, sk+1)f(k+2)(−i) ⊗ g(k+2)i(dxk+2|hk, xk+1, sk+1);
ii. since f(k+1)0 is absolutely continuous with respect to λk+1, the above
inequality also holds for f(k+1)0(hk)-almost all sk+1 ∈ Sk+1 and all xk+1 ∈Xk+1 such that (hk, xk+1, sk+1) ∈ Hk+1.
Repeating the above argument, one can show that∫∏
m≥k+1(Xm×Sm)u(hk, x, s)%(hk,f)(d(x, s))
≥∫∏
m≥k+1(Xm×Sm)u(hk, x, s)%(hk,fM+1)
(d(x, s))
for any M > k. Since∫∏
m≥k+1(Xm×Sm)u(hk, x, s)%(hk,fM+1)
(d(x, s))
converges to ∫∏
m≥k+1(Xm×Sm)u(hk, x, s)%(hk,(gi,f−i))(d(x, s))
as M goes to infinity, we have∫∏
m≥k+1(Xm×Sm)u(hk, x, s)%(hk,f)(d(x, s))
≥∫∏
m≥k+1(Xm×Sm)u(hk, x, s)%(hk,(gi,f−i))(d(x, s)).
Therefore, {fki}i∈I is a subgame-perfect equilibrium.
58
By Lemma 20 and Proposition 4, the correspondence Φ(Q∞t+1) is measurable,
nonempty and compact valued. By Lemma 3 (3), it has a measurable selection.
Then Theorem 1 follows from the above lemma.
For t ≥ 1 and ht−1 ∈ Ht−1, recall that Et(ht−1) is the set of payoff vectors of
subgame-perfect equilibria in the subgame ht−1. The following lemma shows that
Et(ht−1) is essentially the same as Q∞t (ht−1).
Lemma 22. For any t ≥ 1, Et(ht−1) = Q∞t (ht−1) for λt−1-almost all ht−1 ∈ Ht−1.
Proof. (1) We will first prove the following claim: for any t and τ , if Et+1(ht) ⊆Qτt+1(ht) for λt-almost all ht ∈ Ht, then Et(ht−1) ⊆ Qτt (ht−1) for λt−1-almost all
ht−1 ∈ Ht−1. We only need to consider the case that t ≤ τ .
By the construction of Φ(Qτt+1) in Subsection 5.3.1, there exists a measurable
subset St−1 ⊆ St−1 with λt−1(St−1) = 1 such that for any ct and ht−1 =
(xt−1, st−1) ∈ Ht−1 with st−1 ∈ St−1, if
1. ct =∫At(ht−1)
∫Stqt+1(ht−1, xt, st)ft0(dst|ht−1)α(dxt), where qt+1(ht−1, ·) is
measurable and qt+1(ht−1, xt, st) ∈ Qτt+1(ht−1, xt, st) for λt-almost all st ∈ Stand xt ∈ At(ht−1);
2. α ∈ ⊗i∈IM(Ati(ht−1)) is a Nash equilibrium in the subgame ht−1 with payoff∫Stqt+1(ht−1, ·, st)ft0(dst|ht−1) and action space
∏i∈I Ati(ht−1),
then ct ∈ Φ(Qτt+1)(ht−1).
Fix a subgame ht−1 = (xt−1, st−1) such that st−1 ∈ St−1. Pick a point
ct ∈ Et(st−1). There exists a strategy profile f such that f is a subgame-perfect
equilibrium in the subgame ht−1 and the payoff is ct. Let ct+1(ht−1, xt, st) be the
payoff vector induced by {fti}i∈I in the subgame (ht, xt, st) ∈ Gr(At) × St. Then
we have
1. ct =∫At(ht−1)
∫Stct+1(ht−1, xt, st)ft0(dst|ht−1)ft(dxt|ht−1);
2. ft(·|ht−1) is a Nash equilibrium in the subgame ht−1 with action space
At(ht−1) and payoff∫Stct+1(ht−1, ·, st)ft0(dst|ht−1).
Since f is a subgame-perfect equilibrium in the subgame ht−1, ct+1(ht−1, xt, st) ∈Et+1(ht−1, xt, st) ⊆ Qτt+1(ht−1, xt, st) for λt-almost all st ∈ St and xt ∈ At(ht−1),
which implies that ct ∈ Φ(Qτt+1)(ht−1) = Qτt (ht−1).
Therefore, Et(ht−1) ⊆ Qτt (ht−1) for λt−1-almost all ht−1 ∈ Ht−1.
(2) For any t > τ , Et ⊆ Qτt . If t ≤ τ , we can start with Eτ+1 ⊆ Qττ+1 and repeat
the argument in (1), then we can show that Et(ht−1) ⊆ Qτt (ht−1) for λt−1-almost
all ht−1 ∈ Ht−1. Thus, Et(ht−1) ⊆ Q∞t (ht−1) for λt−1-almost all ht−1 ∈ Ht−1.
(3) Suppose that ct is a measurable selection from Φ(Q∞t+1). Apply Proposi-
tion 5 recursively to obtain Borel measurable mappings {fki}i∈I for k ≥ t. By
59
Lemma 21, ct(ht−1) is a subgame-perfect equilibrium payoff vector for λt−1-almost
all ht−1 ∈ Ht−1. Consequently, Φ(Q∞t+1)(ht−1) ⊆ Et(ht−1) for λt−1-almost all
ht−1 ∈ Ht−1.
By Lemma 20, Et(ht−1) = Q∞t (ht−1) = Φ(Q∞t+1)(ht−1) for λt−1-almost all
ht−1 ∈ Ht−1.
Therefore, we have proved Proposition 1.
5.4 Proof of Proposition 2
We will highlight the needed changes in comparison with the proofs presented in
Subsections 5.3.1-5.3.3.
1. Backward induction. We first consider stage t with Nt = 1.
If Nt = 1, then St = {st}. Thus, Pt(ht−1, xt) = Qt+1(ht−1, xt, st), which is
nonempty and compact valued, and essentially sectionally upper hemicontinuous
on Xt × St−1. Notice that Pt may not be convex valued.
We first assume that Pt is upper hemicontinuous. Suppose that j is the player
who is active in this period. Consider the correspondence Φt : Ht−1 → Rn ×M(Xt)×4(Xt) defined as follows: (v, α, µ) ∈ Φt(ht−1) if
1. v = pt(ht−1, At(−j)(ht−1), x∗tj) such that pt(ht−1, ·) is a measurable selection
of Pt(ht−1, ·);35
2. x∗tj ∈ Atj(ht−1) is a maximization point of player j given the payoff function
ptj(ht−1, At(−j)(ht−1), ·) and the action space Atj(ht−1), αi = δAti(ht−1) for
i 6= j and αj = δx∗tj ;
3. µ = δpt(ht−1,At(−j)(ht−1),x∗tj).
This is a single agent problem. We need to show that Φt is nonempty and compact
valued, and upper hemicontinuous.
If Pt is nonempty, convex and compact valued, and upper hemicontinuous,
then we can use Lemma 16, the main result of Simon and Zame (1990), to prove
the nonemptiness, compactness, and upper hemicontinuity of Φt. In Simon and
Zame (1990), the only step they need the convexity of Pt for the proof of their main
theorem is Lemma 2 therein. However, the one-player pure-strategy version of their
Lemma 2, stated in the following, directly follows from the upper hemicontinuity
of Pt without requiring the convexity.
Let Z be a compact metric space, and {zn}n≥0 ⊆ Z. Let P : Z → R+ be
a bounded, upper hemicontinuous correspondence with nonempty and
compact values. For each n ≥ 1, let qn be a Borel measurable selection
35Note that At(−j) is point valued since all players other than j are inactive.
60
of P such that qn(zn) = dn. If zn converges to z0 and dn converges to
some d0, then d0 ∈ P (z0).
Repeat the argument in the proof of the main theorem of Simon and Zame
(1990), one can show that Φt is nonempty and compact valued, and upper
hemicontinuous.
Then we go back to the case that Pt is nonempty and compact valued,
and essentially sectionally upper hemicontinuous on Xt × St−1. Recall that we
proved Proposition 4 based on Lemma 16. If Pt is essentially sectionally upper
hemicontinuous on Xt× St−1, we can show the following result based on a similar
argument as in Subsections 5.2: there exists a bounded, measurable, nonempty
and compact valued correspondence Φt from Ht−1 to Rn ×M(Xt) ×4(Xt) such
that Φt is essentially sectionally upper hemicontinuous on Xt−1 × St−1, and for
λt−1-almost all ht−1 ∈ Ht−1, (v, α, µ) ∈ Φt(ht−1) if
1. v = pt(ht−1, At(−j)(ht−1), x∗tj) such that pt(ht−1, ·) is a measurable selection
of Pt(ht−1, ·);
2. x∗tj ∈ Atj(ht−1) is a maximization point of player j given the payoff function
ptj(ht−1, At(−j)(ht−1), ·) and the action space Atj(ht−1), αi = δAti(ht−1) for
i 6= j and αj = δx∗tj ;
3. µ = δpt(ht−1,At(−j)(ht−1),x∗tj).
Next we consider the case that Nt = 0. Suppose that the correspondence
Qt+1 from Ht to Rn is bounded, measurable, nonempty and compact valued, and
essentially sectionally upper hemicontinuous on Xt × St. For any (ht−1, xt, st) ∈Gr(At), let
Rt(ht−1, xt, st) =
∫St
Qt+1(ht−1, xt, st, st)ft0(dst|ht−1, xt, st)
=
∫St
Qt+1(ht−1, xt, st, st)ϕt0(ht−1, xt, st, st)λt(dst).
Then following the same argument as in Subsection 5.3.1, one can show that
Rt is a nonempty, convex and compact valued, and essentially sectionally upper
hemicontinuous correspondence on Xt × St.For any ht−1 ∈ Ht−1 and xt ∈ At(ht−1), let
Pt(ht−1, xt) =
∫At0(ht−1,xt)
Rt(ht−1, xt, st)ft0(dst|ht−1, xt).
By Lemma 8, Pt is nonempty, convex and compact valued, and essentially
sectionally upper hemicontinuous on Xt × St−1. The rest of the step remains
the same as in Subsection 5.3.1.
61
2. Forward induction: unchanged.
3. Infinite horizon: we need to slightly modify the definition of Ξm1t for any
m1 ≥ t ≥ 1. Fix any t ≥ 1. Define a correspondence Ξtt as follows: in the subgame
ht−1,
Ξtt(ht−1) = (M(At(ht−1)) � ft0(ht−1, ·))⊗ λt.
For any m1 > t, suppose that the correspondence Ξm1−1t has been defined. Then we
can define a correspondence Ξm1t : Ht−1 →M
(∏t≤m≤m1
(Xm × Sm))
as follows:
Ξm1t (ht−1) =
{g(ht−1) �
((ξm1(ht−1, ·) � fm10(ht−1, ·))⊗ λm1
):
g is a Borel measurable selection of Ξm1−1t ,
ξm1 is a Borel measurable selection of M(Am1)}.
Then the result in Subsection 5.3.3 is true with the above Ξm1t .
Consequently, a subgame-perfect equilibrium exists.
5.5 Proof of Proposition 3
We will describe the necessary changes in comparison with the proofs presented in
Subsections 5.3.1-5.3.3 and 5.4.
1. Backward induction. For any t ≥ 1, suppose that the correspondence
Qt+1 from Ht to Rn is bounded, nonempty and compact valued, and upper
hemicontinuous on Xt.
If Nt = 1, then St = {st}. Thus, Pt(ht−1, xt) = Qt+1(ht−1, xt, st), which
is nonempty and compact valued, and upper hemicontinuous. Then define the
correspondence Φt from Ht−1 to Rn ×M(Xt)×4(Xt) as (v, α, µ) ∈ Φt(ht−1) if
1. v = pt(ht−1, At(−j)(ht−1), x∗tj) such that pt(ht−1, ·) is a measurable selection
of Pt(ht−1, ·);
2. x∗tj ∈ Atj(ht−1) is a maximization point of player j given the payoff function
ptj(ht−1, At(−j)(ht−1), ·) and the action space Atj(ht−1), αi = δAti(ht−1) for
i 6= j and αj = δx∗tj ;
3. µ = δpt(ht−1,At(−j)(ht−1),x∗tj).
As discussed in Subsection 5.4, Φt is nonempty and compact valued, and upper
hemicontinuous.
62
When Nt = 0, for any ht−1 ∈ Ht−1 and xt ∈ At(ht−1),
Pt(ht−1, xt) =
∫At0(ht−1,xt)
Qt+1(ht−1, xt, st)ft0(dst|ht−1, xt).
Let coQt+1(ht−1, xt, st) be the convex hull of Qt+1(ht−1, xt, st). Because Qt+1 is
bounded, nonempty and compact valued, coQt+1 is bounded, nonempty, convex
and compact valued. By Lemma 3 (8), coQt+1 is upper hemicontinuous.
Notice that ft0(·|ht−1, xt) is atomless and Qt+1 is nonempty and compact
valued. By Lemma 5,
Pt(ht−1, xt) =
∫At0(ht−1,xt)
coQt+1(ht−1, xt, st)ft0(dst|ht−1, xt).
By Lemma 8, Pt is bounded, nonempty, convex and compact valued, and upper
hemicontinuous. Then instead of relying on Proposition 4, we now use Lemma 16
to conclude that Φt is bounded, nonempty and compact valued, and upper
hemicontinuous.
2. Forward induction. The first step is much simpler.
For any {(hkt−1, vk)}1≤k≤∞ ⊆ Gr(Φt(Qt+1)) such that (hkt−1, v
k) converges to
(h∞t−1, v∞), pick (αk, µk) such that (vk, αk, µk) ∈ Φt(h
kt−1) for 1 ≤ k < ∞. Since
Φt is upper hemicontinuous and compact valued, there exists a subsequence of
(vk, αk, µk), say itself, such that (vk, αk, µk) converges to some (v∞, α∞, µ∞) ∈Φt(h
∞t−1) due to Lemma 2 (5). Thus, (α∞, µ∞) ∈ Ψt(h
∞t−1, v
∞), which implies that
Ψt is also upper hemicontinuous and compact valued. By Lemma 3 (3), Ψt has a
Borel measurable selection ψt. Given a Borel measurable selection qt of Φ(Qt+1),
one can let φt(ht−1) = (qt(ht−1), ψt(ht−1, qt(ht−1))). Then φt is a Borel measurable
selection of Φt.
Steps 2 and 3 are unchanged.
3. Infinite horizon. We do not need to consider Ξm1t for any m1 ≥ t ≥ 1.
Instead of relying on Lemma 14, we can use Lemma 10 (3) to prove Lemma 18. The
proof of Lemma 19 is much simpler. Notice that the boundedness, nonemptiness,
compactness and upper hemicontinuity of Qτt for the case t > τ is immediate. Then
one can apply the backward induction as in Lemma 19 to show the corresponding
properties of Qτt for the case t ≤ τ . Following the same arguments, one can show
that Lemmas 20-22 now hold for all ht−1 ∈ Ht−1 and all t ≥ 1.
63
References
D. Abreu, D. Pearce and E. Stacchetti, Toward a Theory of Discounted RepeatedGames with Imperfect Monitoring, Econometrica 58 (1990), 1041–1063.
C. D. Aliprantis and K. C. Border, Infinite Dimensional Analysis: A Hitchhiker’sGuide, Springer, Berlin, 2006.
G. S. Becker and K. M. Murphy, A Theory of Rational Addiction, Journal ofPolitical Economy 96 (1988), 675–700.
V. I. Bogachev, Measure Theory, Volume II, Springer-Verlag, Berlin Heidelberg,2007.
T. Borgers, Perfect Equilibrium Histories of Finite and Infinite Horizon Games,Journal of Economic Theory 47 (1989), 218–227.
T. Borgers, Upper Hemicontinuity of The Correspondence of Subgame-PerfectEquilibrium Outcomes, Journal of Mathematical Economics 20 (1991), 89–106.
L. Brown and B. M. Schreiber, Approximation and Extension of RandomFunctions, Monatshefte fur Mathematik 107 (1989), 111–123.
C. Castaing, P. R. De Fitte and M. Valadier, Young Measures on TopologicalSpaces: with Applications in Control Theory and Probability Theory, Vol. 571,Springer Science & Business Media, 2004.
C. Fershtman and M. I. Kamien, Dynamic Duopolistic Competition with StickyPrices, Econometrica 55 (1987), 1151–1164.
D. Fudenberg and D. Levine, Subgame-Perfect Equilibria of Finite and Infinite-Horizon Games, Journal of Economic Theory 31 (1983), 251–268.
D. Fudenberg and J. Tirole, Game Theory, MIT Press, Cambridge, 1991.
S. Goldman, Consistent Plans, Review of Economic Studies 47 (1980), 533–537.
E. J. Green and R. H. Porter, Noncooperative Collusion under Imperfect PriceInformation, Econometrica 52 (1984), 87–100.
C. Harris, Existence and Characterization of Perfect Equilibrium in Games ofPerfect Information, Econometrica 53 (1985), 613–628.
C. Harris, The Existence of Subgame-Perfect Equilibrium with and without MarkovStrategies: A Case for Extensive-Form Correlation, Discussion Paper No. 53,Nuffield College, Oxford, 1990.
C. Harris, P. J. Reny and A. Robson, The Existence of Subgame-PerfectEquilibrium in Continuous Games with Almost Perfect Information: A Casefor Public Randomization, Econometrica 63 (1995), 507–544.
64
M. Hellwig and W. Leininger, On the Existence of Subgame-Perfect Equilibriumin Infinite-Action Games of Perfect Information, Journal of Economic Theory43 (1987), 55–75.
M. Hellwig, W. Leininger, P. J. Reny and A. J. Robson, Subgame-PerfectEquilibrium in Continuous Games of Perfect Information: An ElementaryApproach to Existence and Approximation by Discrete Games, Journal ofEconomic Theory 52 (1990), 406–422.
W. Hildenbrand, Core and Equilibria of a Large Economy, Princeton UniversityPress, Princeton, NJ, 1974.
A. Kucia, Some Results on Caratheodory Selections and Extensions, Journal ofMathematical Analysis and Applications 223 (1998), 302–318.
E. G. J. Luttmer and T. Mariotti, The Existence of Subgame-Perfect Equilibriumin Continuous Games with Almost Perfect Information: A Comment,Econometrica 71 (2003), 1909–1911.
T. Mariotti, Subgame-Perfect Equilibrium Outcomes in Continuous Games ofAlmost Perfect Information, Journal of Mathematical Economics 34 (2000),99–128.
J.-F. Mertens, A Measurable ‘Measurable Choice’ Theorem, in Proceedings of theNATO Advanced Study Institute on Stochastic Games and Applications, ed. byA. Neyman and S. Sorin. NATO Science Series. Dordrecht: Kluwer Academic,2003, 107–130.
J.-F. Mertens and T. Parthasarathy, Equilibria for Discounted Stochastic Games,in Proceedings of the NATO Advanced Study Institute on Stochastic Games andApplications, ed. by A. Neyman and S. Sorin. NATO Science Series. Dordrecht:Kluwer Academic, 2003, 131–172.
E. Michael, A Selection Theorem, Proceedings of the American MathematicalSociety 17 (1966), 1404–1406.
R. B. Myerson and P. J. Reny, Sequential Equilibria of Multi-stage Games withInfinite Sets of Types and Actions, working paper, University of Chicago, 2015.
B. Peleg and M. Yaari, On the Existence of a Consistent Course of Action WhenTastes Are Changing, Review of Economic Studies 40 (1973), 391–401.
E. Phelps and R. Pollak, On Second Best National Savings and Game EquilibriumGrowth, Review of Economic Studies 35 (1968), 185–199.
P. J. Reny and A. J. Robson, Existence of Subgame-Perfect Equilibrium withPublic Randomization: A Short Proof, Economics Bulletin 24 (2002), 1–8.
J. J. Rotemberg and G. Saloner, A Supergame-Theoretic Model of Price Warsduring Booms, American Economic Review 76 (1986), 390–407.
65
H. E. Ryder and G. M. Heal, Optimal Growth with Intertemporally DependentPreferences, Review of Economic Studies 40 (1973), 1–31.
R. Selten, Spieltheoretische Behandlung Eines Oligopolmodells Mit Nachfragetra-heit, Z. Gesamte Staatwissenschaft 12 (1965), 301–324.
L. K. Simon and W. R. Zame, Discontinuous Games and Endogenous SharingRules, Econometrica 58 (1990), 861–872.
Y. N. Sun, Integration of Correspondences on Loeb Spaces, Transactions of theAmerican Mathematical Society 349 (1997), 129–153.
66